Ball An Elementary Introduction to Modern Convex Geometry

background image

Flavors of Geometry

MSRI Publications

Volume 31, 1997

An Elementary Introduction

to Modern Convex Geometry

KEITH BALL

Contents

Preface

1

Lecture 1. Basic Notions

2

Lecture 2. Spherical Sections of the Cube

8

Lecture 3. Fritz John’s Theorem

13

Lecture 4. Volume Ratios and Spherical Sections of the Octahedron

19

Lecture 5. The Brunn–Minkowski Inequality and Its Extensions

25

Lecture 6. Convolutions and Volume Ratios: The Reverse Isoperimetric

Problem

32

Lecture 7. The Central Limit Theorem and Large Deviation Inequalities 37
Lecture 8. Concentration of Measure in Geometry

41

Lecture 9. Dvoretzky’s Theorem

47

Acknowledgements

53

References

53

Index

55

Preface

These notes are based, somewhat loosely, on three series of lectures given by

myself, J. Lindenstrauss and G. Schechtman, during the Introductory Workshop
in Convex Geometry held at the Mathematical Sciences Research Institute in
Berkeley, early in 1996. A fourth series was given by B. Bollob´

as, on rapid

mixing and random volume algorithms; they are found elsewhere in this book.

The material discussed in these notes is not, for the most part, very new, but

the presentation has been strongly influenced by recent developments: among
other things, it has been possible to simplify many of the arguments in the light
of later discoveries. Instead of giving a comprehensive description of the state of
the art, I have tried to describe two or three of the more important ideas that
have shaped the modern view of convex geometry, and to make them as accessible

1

background image

2

KEITH BALL

as possible to a broad audience. In most places, I have adopted an informal style
that I hope retains some of the spontaneity of the original lectures. Needless to
say, my fellow lecturers cannot be held responsible for any shortcomings of this
presentation.

I should mention that there are large areas of research that fall under the

very general name of convex geometry, but that will barely be touched upon in
these notes. The most obvious such area is the classical or “Brunn–Minkowski”
theory, which is well covered in [Schneider 1993]. Another noticeable omission is
the combinatorial theory of polytopes: a standard reference here is [Brøndsted
1983].

Lecture 1. Basic Notions

The topic of these notes is convex geometry. The objects of study are con-

vex bodies: compact, convex subsets of Euclidean spaces, that have nonempty
interior. Convex sets occur naturally in many areas of mathematics: linear pro-
gramming, probability theory, functional analysis, partial differential equations,
information theory, and the geometry of numbers, to name a few.

Although convexity is a simple property to formulate, convex bodies possess

a surprisingly rich structure. There are several themes running through these
notes: perhaps the most obvious one can be summed up in the sentence: “All
convex bodies behave a bit like Euclidean balls.” Before we look at some ways in
which this is true it is a good idea to point out ways in which it definitely is not.
This lecture will be devoted to the introduction of a few basic examples that we
need to keep at the backs of our minds, and one or two well known principles.

The only notational conventions that are worth specifying at this point are

the following. We will use

| · | to denote the standard Euclidean norm on

R

n

. For

a body K, vol(K) will mean the volume measure of the appropriate dimension.

The most fundamental principle in convexity is the Hahn–Banach separation

theorem, which guarantees that each convex body is an intersection of half-spaces,
and that at each point of the boundary of a convex body, there is at least one
supporting hyperplane. More generally, if K and L are disjoint, compact, convex
subsets of

R

n

, then there is a linear functional φ :

R

n

R

for which φ(x) < φ(y)

whenever x

∈ K and y ∈ L.

The simplest example of a convex body in

R

n

is the cube, [

1, 1]

n

. This does

not look much like the Euclidean ball. The largest ball inside the cube has radius
1, while the smallest ball containing it has radius

n, since the corners of the

cube are this far from the origin. So, as the dimension grows, the cube resembles
a ball less and less.

The second example to which we shall refer is the n-dimensional regular solid

simplex: the convex hull of n + 1 equally spaced points. For this body, the ratio
of the radii of inscribed and circumscribed balls is n: even worse than for the
cube. The two-dimensional case is shown in Figure 1. In Lecture 3 we shall see

background image

AN ELEMENTARY INTRODUCTION TO MODERN CONVEX GEOMETRY

3

Figure 1. Inscribed and circumscribed spheres for an n-simplex.

that these ratios are extremal in a certain well-defined sense.

Solid simplices are particular examples of cones. By a cone in

R

n

we just mean

the convex hull of a single point and some convex body of dimension n

1 (Figure

2). In

R

n

, the volume of a cone of height h over a base of (n

1)-dimensional

volume B is Bh/n.

The third example, which we shall investigate more closely in Lecture 4, is the

n-dimensional “octahedron”, or cross-polytope: the convex hull of the 2n points
(

±1, 0, 0, . . . , 0), (0, ±1, 0, . . . , 0), . . . , (0, 0, . . . , 0, ±1). Since this is the unit ball

of the `

1

norm on

R

n

, we shall denote it B

n

1

. The circumscribing sphere of B

n

1

has radius 1, the inscribed sphere has radius 1/

n; so, as for the cube, the ratio

is

n: see Figure 3, left.

B

n

1

is made up of 2

n

pieces similar to the piece whose points have nonnegative

coordinates, illustrated in Figure 3, right. This piece is a cone of height 1 over
a base, which is the analogous piece in

R

n−1

. By induction, its volume is

1

n

·

1

n

1

· · · · ·

1

2

· 1 =

1

n!

,

and hence the volume of B

n

1

is 2

n

/n!.

Figure 2. A cone.

background image

4

KEITH BALL

(1, 0, . . . , 0)

(

1

n

, . . . ,

1

n

)

(1, 0, 0)

(0, 1, 0)

(0, 0, 1)

Figure 3. The cross-polytope (left) and one orthant thereof (right).

The final example is the Euclidean ball itself,

B

n

2

=



x

R

n

:

n

X

1

x

2

i

1



.

We shall need to know the volume of the ball: call it v

n

. We can calculate the

surface “area” of B

n

2

very easily in terms of v

n

: the argument goes back to the

ancients. We think of the ball as being built of thin cones of height 1: see Figure
4, left. Since the volume of each of these cones is 1/n times its base area, the
surface of the ball has area nv

n

. The sphere of radius 1, which is the surface of

the ball, we shall denote S

n−1

.

To calculate v

n

, we use integration in spherical polar coordinates. To specify

a point x we use two coordinates: r, its distance from 0, and θ, a point on the
sphere, which specifies the direction of x. The point θ plays the role of n

1 real

coordinates. Clearly, in this representation, x = : see Figure 4, right. We can

x

θ

0

Figure 4. Computing the volume of the Euclidean ball.

background image

AN ELEMENTARY INTRODUCTION TO MODERN CONVEX GEOMETRY

5

write the integral of a function on

R

n

as

Z

R

n

f =

Z

r=0

Z

S

n−1

f () “r

n−1

dr.

(1.1)

The factor r

n−1

appears because the sphere of radius r has area r

n−1

times that

of S

n−1

. The notation “” stands for “area” measure on the sphere: its total

mass is the surface area nv

n

. The most important feature of this measure is

its rotational invariance: if A is a subset of the sphere and U is an orthogonal
transformation of

R

n

, then U A has the same measure as A. Throughout these

lectures we shall normalise integrals like that in (1.1) by pulling out the factor
nv

n

, and write

Z

R

n

f = nv

n

Z

0

Z

S

n−1

f ()r

n−1

(θ) dr

where σ = σ

n−1

is the rotation-invariant measure on S

n−1

of total mass 1. To

find v

n

, we integrate the function

x

7→ exp



1

2

n

X

1

x

2

i



both ways. This function is at once invariant under rotations and a product of
functions depending upon separate coordinates; this is what makes the method
work. The integral is

Z

R

n

f =

Z

R

n

n

Y

1

e

−x

2

i

/2

dx =

n

Y

1

Z

−∞

e

−x

2

i

/2

dx

i



=

2π



n

.

But this equals

nv

n

Z

0

Z

S

n−1

e

−r

2

/2

r

n−1

dσ dr = nv

n

Z

0

e

−r

2

/2

r

n−1

dr = v

n

2

n/2

Γ



n

2

+ 1



.

Hence

v

n

=

π

n/2

Γ

n

2

+ 1

.

This is extremely small if n is large. From Stirling’s formula we know that

Γ



n

2

+ 1



2π e

−n/2



n

2



(n+1)/2

,

so that v

n

is roughly



r

2πe

n



n

.

To put it another way, the Euclidean ball of volume 1 has radius about

r

n

2πe

,

which is pretty big.

background image

6

KEITH BALL

r = v

1/n

n

Figure 5. Comparing the volume of a ball with that of its central slice.

This rather surprising property of the ball in high-dimensional spaces is per-

haps the first hint that our intuition might lead us astray. The next hint is
provided by an answer to the following rather vague question: how is the mass
of the ball distributed? To begin with, let’s estimate the (n

1)-dimensional

volume of a slice through the centre of the ball of volume 1. The ball has radius

r = v

1/n

n

(Figure 5). The slice is an (n

1)-dimensional ball of this radius, so its volume is

v

n−1

r

n−1

= v

n−1



1

v

n



(n−1)/n

.

By Stirling’s formula again, we find that the slice has volume about

e when

n is large. What are the (n

1)-dimensional volumes of parallel slices? The

slice at distance x from the centre is an (n

1)-dimensional ball whose radius is

r

2

− x

2

(whereas the central slice had radius r), so the volume of the smaller

slice is about

e



r

2

− x

2

r



n−1

=

e



1

x

2

r

2



(n−1)/2

.

Since r is roughly

p

n/(2πe), this is about

e



1

2πex

2

n



(n−1)/2

e exp(

−πex

2

).

Thus, if we project the mass distribution of the ball of volume 1 onto a single

direction, we get a distribution that is approximately Gaussian (normal) with
variance 1/(2πe). What is remarkable about this is that the variance does not
depend upon n. Despite the fact that the radius of the ball of volume 1 grows
like

p

n/(2πe), almost all of this volume stays within a slab of fixed width: for

example, about 96% of the volume lies in the slab

{x ∈

R

n

:

1

2

≤ x

1

1

2

}.

See Figure 6.

background image

AN ELEMENTARY INTRODUCTION TO MODERN CONVEX GEOMETRY

7

n = 2

n = 16

n = 120

96%

Figure 6. Balls in various dimensions, and the slab that contains about 96% of
each of them.

So the volume of the ball concentrates close to any subspace of dimension

n

1. This would seem to suggest that the volume concentrates near the centre

of the ball, where the subspaces all meet. But, on the contrary, it is easy to see
that, if n is large, most of the volume of the ball lies near its surface. In objects of
high dimension, measure tends to concentrate in places that our low-dimensional
intuition considers small. A considerable extension of this curious phenomenon
will be exploited in Lectures 8 and 9.

To finish this lecture, let’s write down a formula for the volume of a general

body in spherical polar coordinates. Let K be such a body with 0 in its interior,
and for each direction θ

∈ S

n−1

let r(θ) be the radius of K in this direction.

Then the volume of K is

nv

n

Z

S

n−1

Z

r(θ)

0

s

n−1

ds dσ = v

n

Z

S

n−1

r(θ)

n

(θ).

This tells us a bit about particular bodies. For example, if K is the cube [

1, 1]

n

,

whose volume is 2

n

, the radius satisfies

Z

S

n−1

r(θ)

n

=

2

n

v

n



r

2n

πe



n

.

So the “average” radius of the cube is about

r

2n

πe

.

This indicates that the volume of the cube tends to lie in its corners, where the
radius is close to

n, not in the middle of its facets, where the radius is close

to 1. In Lecture 4 we shall see that the reverse happens for B

n

1

and that this has

a surprising consequence.

background image

8

KEITH BALL

If K is (centrally) symmetric, that is, if

−x ∈ K whenever x ∈ K, then K is

the unit ball of some norm

k · k

K

on

R

n

:

K =

{x : kxk

K

1} .

This was already mentioned for the octahedron, which is the unit ball of the `

1

norm

kxk =

n

X

1

|x

i

|.

The norm and radius are easily seen to be related by

r(θ) =

1

kθk

,

for θ

∈ S

n−1

,

since r(θ) is the largest number r for which

∈ K. Thus, for a general sym-

metric body K with associated norm

k · k, we have this formula for the volume:

vol(K) = v

n

Z

S

n−1

kθk

−n

(θ).

Lecture 2. Spherical Sections of the Cube

In the first lecture it was explained that the cube is rather unlike a Euclidean

ball in

R

n

: the cube [

1, 1]

n

includes a ball of radius 1 and no more, and is

included in a ball of radius

n and no less. The cube is a bad approximation

to the Euclidean ball. In this lecture we shall take this point a bit further. A
body like the cube, which is bounded by a finite number of flat facets, is called a
polytope. Among symmetric polytopes, the cube has the fewest possible facets,
namely 2n. The question we shall address here is this:

If K is a polytope in

R

n

with m facets, how well can K approximate the

Euclidean ball?

Let’s begin by clarifying the notion of approximation. To simplify matters we
shall only consider symmetric bodies. By analogy with the remarks above, we
could define the distance between two convex bodies K and L to be the smallest
d for which there is a scaled copy of L inside K and another copy of L, d times
as large, containing K. However, for most purposes, it will be more convenient
to have an affine-invariant notion of distance: for example we want to regard all
parallelograms as the same. Therefore:

Definition.

The distance d(K, L) between symmetric convex bodies K and

L is the least positive d for which there is a linear image ˜

L of L such that

˜

L

⊂ K ⊂ d˜L. (See Figure 7.)

Note that this distance is multiplicative, not additive: in order to get a metric (on
the set of linear equivalence classes of symmetric convex bodies) we would need to
take log d instead of d. In particular, if K and L are identical then d(K, L) = 1.

background image

AN ELEMENTARY INTRODUCTION TO MODERN CONVEX GEOMETRY

9

˜

L

d ˜

L

K

Figure 7. Defining the distance between K and L.

Our observations of the last lecture show that the distance between the cube
and the Euclidean ball in

R

n

is at most

n. It is intuitively clear that it really

is

n, i.e., that we cannot find a linear image of the ball that sandwiches the

cube any better than the obvious one. A formal proof will be immediate after
the next lecture.

The main result of this lecture will imply that, if a polytope is to have small

distance from the Euclidean ball, it must have very many facets: exponentially
many in the dimension n.

Theorem 2.1.

Let K be a (symmetric) polytope in

R

n

with d(K, B

n

2

) = d. Then

K has at least e

n/(2d

2

)

facets. On the other hand , for each n, there is a polytope

with 4

n

facets whose distance from the ball is at most 2.

The arguments in this lecture, including the result just stated, go back to the
early days of packing and covering problems. A classical reference for the subject
is [Rogers 1964].

Before we embark upon a proof of Theorem 2.1, let’s look at a reformulation

that will motivate several more sophisticated results later on.

A symmetric

convex body in

R

n

with m pairs of facets can be realised as an n-dimensional

slice (through the centre) of the cube in

R

m

. This is because such a body is

the intersection of m slabs in

R

n

, each of the form

{x : |hx, vi| ≤ 1}, for some

nonzero vector v in

R

n

. This is shown in Figure 8.

Thus K is the set

{x : |hx, v

i

i| ≤ 1 for 1 ≤ i ≤ m}, for some sequence (v

i

)

m

1

of

vectors in

R

n

. The linear map

T : x

7→ (hx, v

1

i, . . . , hx, v

m

i)

embeds

R

n

as a subspace H of

R

m

, and the intersection of H with the cube

[

1, 1]

m

is the set of points y in H for which

|y

i

| ≤ 1 for each coordinate i. So

this intersection is the image of K under T . Conversely, any n-dimensional slice
of [

1, 1]

m

is a body with at most m pairs of faces. Thus, the result we are

aiming at can be formulated as follows:

The cube in

R

m

has almost spherical sections whose dimension n is roughly

log m and not more.

background image

10

KEITH BALL

Figure 8. Any symmetric polytope is a section of a cube.

In Lecture 9 we shall see that all symmetric m-dimensional convex bodies have
almost spherical sections of dimension about log m. As one might expect, this is
a great deal more difficult to prove for general bodies than just for the cube.

For the proof of Theorem 2.1, let’s forget the symmetry assumption again and

just ask for a polytope

K =

{x : hx, v

i

i ≤ 1 for 1 ≤ i ≤ m}

with m facets for which

B

n

2

⊂ K ⊂ dB

n

2

.

What do these inclusions say about the vectors (v

i

)? The first implies that each

v

i

has length at most 1, since, if not, v

i

/

|v

i

| would be a vector in B

n

2

but not

in K. The second inclusion says that if x does not belong to dB

n

2

then x does

not belong to K: that is, if

|x| > d, there is an i for which hx, v

i

i > 1. This is

equivalent to the assertion that for every unit vector θ there is an i for which

hθ, v

i

i ≥

1

d

.

Thus our problem is to look for as few vectors as possible, v

1

, v

2

, . . . , v

m

, of

length at most 1, with the property that for every unit vector θ there is some v

i

with

hθ, v

i

i ≥ 1/d. It is clear that we cannot do better than to look for vectors

of length exactly 1: that is, that we may assume that all facets of our polytope
touch the ball. Henceforth we shall discuss only such vectors.

For a fixed unit vector v and ε

[0, 1), the set C(ε, v) of θ ∈ S

n−1

for which

hθ, vi ≥ ε is called a spherical cap (or just a cap); when we want to be precise,
we will call it the ε-cap about v. (Note that ε does not refer to the radius!) See
Figure 9, left.

We want every θ

∈ S

n−1

to belong to at least one of the (1/d)-caps determined

by the (v

i

). So our task is to estimate the number of caps of a given size needed

to cover the entire sphere. The principal tool for doing this will be upper and
lower estimates for the area of a spherical cap. As in the last lecture, we shall

background image

AN ELEMENTARY INTRODUCTION TO MODERN CONVEX GEOMETRY

11

0

ε

v

v

r

Figure 9. Left: ε-cap C(ε, v) about v. Right: cap of radius r about v.

measure this area as a proportion of the sphere: that is, we shall use σ

n−1

as

our measure. Clearly, if we show that each cap occupies only a small proportion
of the sphere, we can conclude that we need plenty of caps to cover the sphere.
What is slightly more surprising is that once we have shown that spherical caps
are not too small, we will also be able to deduce that we can cover the sphere
without using too many.

In order to state the estimates for the areas of caps, it will sometimes be

convenient to measure the size of a cap in terms of its radius, instead of using
the ε measure. The cap of radius r about v is



θ

∈ S

n−1

:

|θ − v| ≤ r

as illustrated in Figure 9, right. (In defining the radius of a cap in this way we
are implicitly adopting a particular metric on the sphere: the metric induced by
the usual Euclidean norm on

R

n

.) The two estimates we shall use are given in

the following lemmas.

Lemma 2.2 (Upper bound for spherical caps).

For 0

≤ ε < 1, the cap

C(ε, u) on S

n−1

has measure at most e

−nε

2

/2

.

Lemma 2.3 (Lower bound for spherical caps).

For 0

≤ r ≤ 2, a cap of

radius r on S

n−1

has measure at least

1

2

(r/2)

n−1

.

We can now prove Theorem 2.1.

Proof.

Lemma 2.2 implies the first assertion of Theorem 2.1 immediately. If

K is a polytope in

R

n

with m facets and if B

n

2

⊂ K ⊂ dB

n

2

, we can find m caps

C(

1
d

, v

i

) covering S

n−1

. Each covers at most e

−n/(2d

2

)

of the sphere, so

m

exp



n

2d

2



.

To get the second assertion of Theorem 2.1 from Lemma 2.3 we need a little

more argument. It suffices to find m = 4

n

points v

1

, v

2

, . . . , v

m

on the sphere so

that the caps of radius 1 centred at these points cover the sphere: see Figure 10.

Such a set of points is called a 1-net for the sphere: every point of the sphere is

background image

12

KEITH BALL

1

1

2

0

Figure 10. The

1

2

-cap has radius 1.

within distance 1 of some v

i

. Now, if we choose a set of points on the sphere any

two of which are at least distance 1 apart, this set cannot have too many points.
(Such a set is called 1-separated.) The reason is that the caps of radius

1

2

about

these points will be disjoint, so the sum of their areas will be at most 1. A cap of
radius

1

2

has area at least

1

4



n

, so the number m of these caps satisfies m

4

n

.

This does the job, because a maximal 1-separated set is automatically a 1-net:
if we can’t add any further points that are at least distance 1 from all the points
we have got, it can only be because every point of the sphere is within distance 1
of at least one of our chosen points. So the sphere has a 1-net consisting of only
4

n

points, which is what we wanted to show.



Lemmas 2.2 and 2.3 are routine calculations that can be done in many ways.
We leave Lemma 2.3 to the dedicated reader. Lemma 2.2, which will be quoted
throughout Lectures 8 and 9, is closely related to the Gaussian decay of the
volume of the ball described in the last lecture. At least for smallish ε (which is
the interesting range) it can be proved as follows.

Proof.

The proportion of S

n−1

belonging to the cap C(ε, u) equals the pro-

portion of the solid ball that lies in the “spherical cone” illustrated in Figure 11.

0

ε

1

− ε

2

Figure 11. Estimating the area of a cap.

background image

AN ELEMENTARY INTRODUCTION TO MODERN CONVEX GEOMETRY

13

As is also illustrated, this spherical cone is contained in a ball of radius

1

− ε

2

(if ε

1/

2), so the ratio of its volume to that of the ball is at most

1

− ε

2



n/2

≤ e

−nε

2

/2

.



In Lecture 8 we shall quote the upper estimate for areas of caps repeatedly. We
shall in fact be using yet a third way to measure caps that differs very slightly
from the C(ε, u) description. The reader can easily check that the preceding
argument yields the same estimate e

−nε

2

/2

for this other description.

Lecture 3. Fritz John’s Theorem

In the first lecture we saw that the cube and the cross-polytope lie at distance

at most

n from the Euclidean ball in

R

n

, and that for the simplex, the distance

is at most n. It is intuitively clear that these estimates cannot be improved. In
this lecture we shall describe a strong sense in which this is as bad as things get.
The theorem we shall describe was proved by Fritz John [1948].

John considered ellipsoids inside convex bodies. If (e

j

)

n

1

is an orthonormal

basis of

R

n

and (α

j

) are positive numbers, the ellipsoid

(

x :

n

X

1

hx, e

j

i

2

α

2

j

1

)

has volume v

n

Q

α

j

. John showed that each convex body contains a unique

ellipsoid of largest volume and, more importantly, he characterised it. He showed
that if K is a symmetric convex body in

R

n

and

E is its maximal ellipsoid then

K

n

E.

Hence, after an affine transformation (one taking

E to B

n

2

) we can arrange that

B

n

2

⊂ K ⊂

nB

n

2

.

A nonsymmetric K may require nB

n

2

, like the simplex, rather than

nB

n

2

.

John’s characterisation of the maximal ellipsoid is best expressed after an

affine transformation that takes the maximal ellipsoid to B

n

2

. The theorem states

that B

n

2

is the maximal ellipsoid in K if a certain condition holds—roughly, that

there be plenty of points of contact between the boundary of K and that of the
ball. See Figure 12.

Theorem 3.1 (John’s Theorem).

Each convex body K contains an unique

ellipsoid of maximal volume. This ellipsoid is B

n

2

if and only if the following

conditions are satisfied : B

n

2

⊂ K and (for some m) there are Euclidean unit

vectors (u

i

)

m

1

on the boundary of K and positive numbers (c

i

)

m

1

satisfying

X

c

i

u

i

= 0

(3.1)

and

X

c

i

hx, u

i

i

2

=

|x|

2

for each x

R

n

.

(3.2)

background image

14

KEITH BALL

Figure 12. The maximal ellipsoid touches the boundary at many points.

According to the theorem the points at which the sphere touches ∂K can be

given a mass distribution whose centre of mass is the origin and whose inertia
tensor is the identity matrix. Let’s see where these conditions come from. The
first condition, (3.1), guarantees that the (u

i

) do not all lie “on one side of the

sphere”. If they did, we could move the ball away from these contact points and
blow it up a bit to obtain a larger ball in K. See Figure 13.

The second condition, (3.2), shows that the (u

i

) behave rather like an or-

thonormal basis in that we can resolve the Euclidean norm as a (weighted) sum
of squares of inner products. Condition (3.2) is equivalent to the statement that

x =

X

c

i

hx, u

i

iu

i

for all x.

This guarantees that the (u

i

) do not all lie close to a proper subspace of

R

n

. If

they did, we could shrink the ball a little in this subspace and expand it in an
orthogonal direction, to obtain a larger ellipsoid inside K. See Figure 14.

Condition (3.2) is often written in matrix (or operator) notation as

X

c

i

u

i

⊗ u

i

= I

n

(3.3)

where I

n

is the identity map on

R

n

and, for any unit vector u, u

⊗ u is the

rank-one orthogonal projection onto the span of u, that is, the map x

7→ hx, uiu.

The trace of such an orthogonal projection is 1. By equating the traces of the

Figure 13. An ellipsoid where all contacts are on one side can grow.

background image

AN ELEMENTARY INTRODUCTION TO MODERN CONVEX GEOMETRY

15

Figure 14. An ellipsoid (solid circle) whose contact points are all near one plane
can grow.

matrices in the preceding equation, we obtain

X

c

i

= n.

In the case of a symmetric convex body, condition (3.1) is redundant, since

we can take any sequence (u

i

) of contact points satisfying condition (3.2) and

replace each u

i

by the pair

±u

i

each with half the weight of the original.

Let’s look at a few concrete examples. The simplest is the cube. For this

body the maximal ellipsoid is B

n

2

, as one would expect. The contact points are

the standard basis vectors (e

1

, e

2

, . . . , e

n

) of

R

n

and their negatives, and they

satisfy

n

X

1

e

i

⊗ e

i

= I

n

.

That is, one can take all the weights c

i

equal to 1 in (3.2). See Figure 15.

The simplest nonsymmetric example is the simplex. In general, there is no

natural way to place a regular simplex in

R

n

, so there is no natural description of

the contact points. Usually the best way to talk about an n-dimensional simplex
is to realise it in

R

n+1

: for example as the convex hull of the standard basis

e

1

e

2

Figure 15. The maximal ellipsoid for the cube.

background image

16

KEITH BALL

Figure 16. K is contained in the convex body determined by the hyperplanes
tangent to the maximal ellipsoid at the contact points.

vectors in

R

n+1

. We leave it as an exercise for the reader to come up with a nice

description of the contact points.

One may get a bit more of a feel for the second condition in John’s Theorem by

interpreting it as a rigidity condition. A sequence of unit vectors (u

i

) satisfying

the condition (for some sequence (c

i

)) has the property that if T is a linear map

of determinant 1, not all the images T u

i

can have Euclidean norm less than 1.

John’s characterisation immediately implies the inclusion mentioned earlier:

if K is symmetric and

E is its maximal ellipsoid then K ⊂

n

E. To check this

we may assume

E = B

n

2

. At each contact point u

i

, the convex bodies B

n

2

and

K have the same supporting hyperplane. For B

n

2

, the supporting hyperplane at

any point u is perpendicular to u. Thus if x

∈ K we have hx, u

i

i ≤ 1 for each i,

and we conclude that K is a subset of the convex body

C =

{x ∈

R

n

:

hx, u

i

i ≤ 1 for 1 ≤ i ≤ m} .

(3.4)

An example of this is shown in Figure 16.

In the symmetric case, the same argument shows that for each x

∈ K we have

|hx, u

i

i| ≤ 1 for each i. Hence, for x ∈ K,

|x|

2

=

X

c

i

hx, u

i

i

2

X

c

i

= n.

So

|x| ≤

n, which is exactly the statement K

n B

n

2

. We leave as a slightly

trickier exercise the estimate

|x| ≤ n in the nonsymmetric case.

Proof of John’s Theorem.

The proof is in two parts, the harder of which

is to show that if B

n

2

is an ellipsoid of largest volume, then we can find an

appropriate system of weights on the contact points. The easier part is to show
that if such a system of weights exists, then B

n

2

is the unique ellipsoid of maximal

volume. We shall describe the proof only in the symmetric case, since the added
complications in the general case add little to the ideas.

background image

AN ELEMENTARY INTRODUCTION TO MODERN CONVEX GEOMETRY

17

We begin with the easier part. Suppose there are unit vectors (u

i

) in ∂K and

numbers (c

i

) satisfying

X

c

i

u

i

⊗ u

i

= I

n

.

Let

E =



x :

n

X

1

hx, e

j

i

2

α

2

j

1



be an ellipsoid in K, for some orthonormal basis (e

j

) and positive α

j

. We want

to show that

n

Y

1

α

j

1

and that the product is equal to 1 only if α

j

= 1 for all j.

Since

E ⊂ K we have that for each i the hyperplane {x : hx, u

i

i = 1} does not

cut

E. This implies that each u

i

belongs to the polar ellipsoid



y :

n

X

1

α

2

j

hy, e

j

i

2

1



.

(The reader unfamiliar with duality is invited to check this.) So, for each i,

n

X

j=1

α

2

j

hu

i

, e

j

i

2

1.

Hence

X

i

c

i

X

j

α

2

j

hu

i

, e

j

i

2

X

c

i

= n.

But the left side of the equality is just

P

j

α

2

j

, because, by condition (3.2), we

have

X

i

c

i

hu

i

, e

j

i

2

=

|e

j

|

2

= 1

for each j. Finally, the fact that the geometric mean does not exceed the arith-
metic mean (the AM/GM inequality) implies that



Y

α

2

j



1/n

1

n

X

α

2

j

1,

and there is equality in the first of these inequalities only if all α

j

are equal to 1.

We now move to the harder direction. Suppose B

n

2

is an ellipsoid of largest

volume in K. We want to show that there is a sequence of contact points (u

i

)

and positive weights (c

i

) with

1

n

I

n

=

1

n

X

c

i

u

i

⊗ u

i

.

We already know that, if this is possible, we must have

X c

i

n

= 1.

background image

18

KEITH BALL

So our aim is to show that the matrix I

n

/n can be written as a convex combina-

tion of (a finite number of) matrices of the form u

⊗ u, where each u is a contact

point. Since the space of matrices is finite-dimensional, the problem is simply
to show that I

n

/n belongs to the convex hull of the set of all such rank-one

matrices,

T =

{u ⊗ u : u is a contact point} .

We shall aim to get a contradiction by showing that if I

n

/n is not in T , we can

perturb the unit ball slightly to get a new ellipsoid in K of larger volume than
the unit ball.

Suppose that I

n

/n is not in T . Apply the separation theorem in the space of

matrices to get a linear functional φ (on this space) with the property that

φ



I

n

n



< φ(u

⊗ u)

(3.5)

for each contact point u. Observe that φ can be represented by an n

× n matrix

H = (h

jk

), so that, for any matrix A = (a

jk

),

φ(A) =

X

jk

h

jk

a

jk

.

Since all the matrices u

⊗ u and I

n

/n are symmetric, we may assume the same

for H. Moreover, since these matrices all have the same trace, namely 1, the
inequality φ(I

n

/n) < φ(u

⊗ u) will remain unchanged if we add a constant to

each diagonal entry of H. So we may assume that the trace of H is 0: but this
says precisely that φ(I

n

) = 0.

Hence, unless the identity has the representation we want, we have found a

symmetric matrix H with zero trace for which

X

jk

h

jk

(u

⊗ u)

jk

> 0

for every contact point u. We shall use this H to build a bigger ellipsoid inside K.

Now, for each vector u, the expression

X

jk

h

jk

(u

⊗ u)

jk

is just the number u

T

Hu. For sufficiently small δ > 0, the set

E

δ

=



x

R

n

: x

T

(I

n

+ δH)x

1

is an ellipsoid and as δ tends to 0 these ellipsoids approach B

n

2

. If u is one of

the original contact points, then

u

T

(I

n

+ δH)u = 1 + δu

T

Hu > 1,

so u does not belong to

E

δ

. Since the boundary of K is compact (and the function

x

7→ x

T

Hx is continuous)

E

δ

will not contain any other point of ∂K as long as

background image

AN ELEMENTARY INTRODUCTION TO MODERN CONVEX GEOMETRY

19

δ is sufficiently small. Thus, for such δ, the ellipsoid

E

δ

is strictly inside K and

some slightly expanded ellipsoid is inside K.

It remains to check that each

E

δ

has volume at least that of B

n

2

. If we denote

by (µ

j

) the eigenvalues of the symmetric matrix I

n

+ δH, the volume of

E

δ

is

v

n

 Q

µ

j

, so the problem is to show that, for each δ, we have

Q

µ

j

1. What

we know is that

P

µ

j

is the trace of I

n

+ δH, which is n, since the trace of H is

0. So the AM/GM inequality again gives

Y

µ

1/n
j

1

n

X

µ

j

1,

as required.



There is an analogue of John’s Theorem that characterises the unique ellipsoid
of minimal volume containing a given convex body. (The characterisation is
almost identical, guaranteeing a resolution of the identity in terms of contact
points of the body and the Euclidean sphere.) This minimal volume ellipsoid
theorem can be deduced directly from John’s Theorem by duality. It follows
that, for example, the ellipsoid of minimal volume containing the cube [

1, 1]

n

is the obvious one: the ball of radius

n. It has been mentioned several times

without proof that the distance of the cube from the Euclidean ball in

R

n

is

exactly

n. We can now see this easily: the ellipsoid of minimal volume outside

the cube has volume

n



n

times that of the ellipsoid of maximal volume inside

the cube. So we cannot sandwich the cube between homothetic ellipsoids unless
the outer one is at least

n times the inner one.

We shall be using John’s Theorem several times in the remaining lectures. At

this point it is worth mentioning important extensions of the result. We can view
John’s Theorem as a description of those linear maps from Euclidean space to a
normed space (whose unit ball is K) that have largest determinant, subject to the
condition that they have norm at most 1: that is, that they map the Euclidean
ball into K. There are many other norms that can be put on the space of linear
maps. John’s Theorem is the starting point for a general theory that builds
ellipsoids related to convex bodies by maximising determinants subject to other
constraints on linear maps. This theory played a crucial role in the development
of convex geometry over the last 15 years. This development is described in
detail in [Tomczak-Jaegermann 1988].

Lecture 4. Volume Ratios and Spherical Sections of the

Octahedron

In the second lecture we saw that the n-dimensional cube has almost spherical

sections of dimension about log n but not more. In this lecture we will examine
the corresponding question for the n-dimensional cross-polytope B

n

1

. In itself,

this body is as far from the Euclidean ball as is the cube in

R

n

: its distance from

the ball, in the sense described in Lecture 2 is

n. Remarkably, however, it has

background image

20

KEITH BALL

almost spherical sections whose dimension is about

1

2

n. We shall deduce this from

what is perhaps an even more surprising statement concerning intersections of
bodies. Recall that B

n

1

contains the Euclidean ball of radius

1

n

. If U is an

orthogonal transformation of

R

n

then U B

n

1

also contains this ball and hence so

does the intersection B

n

1

∩ UB

n

1

. But, whereas B

n

1

does not lie in any Euclidean

ball of radius less than 1, we have the following theorem [Kaˇsin 1977]:

Theorem 4.1.

For each n, there is an orthogonal transformation U for which

the intersection B

n

1

∩ UB

n

1

is contained in the Euclidean ball of radius 32/

n

(and contains the ball of radius 1/

n).

(The constant 32 can easily be improved: the important point is that it is inde-
pendent of the dimension n.) The theorem states that the intersection of just two
copies of the n-dimensional octahedron may be approximately spherical. Notice
that if we tried to approximate the Euclidean ball by intersecting rotated copies
of the cube, we would need exponentially many in the dimension, because the
cube has only 2n facets and our approximation needs exponentially many facets.
In contrast, the octahedron has a much larger number of facets, 2

n

: but of course

we need to do a lot more than just count facets in order to prove Theorem 4.1.
Before going any further we should perhaps remark that the cube has a property
that corresponds to Theorem 4.1. If Q is the cube and U is the same orthogonal
transformation as in the theorem, the convex hull

conv(Q

∪ UQ)

is at distance at most 32 from the Euclidean ball.

In spirit, the argument we shall use to establish Theorem 4.1 is Kaˇsin’s original

one but, following [Szarek 1978], we isolate the main ingredient and we organise
the proof along the lines of [Pisier 1989]. Some motivation may be helpful. The
ellipsoid of maximal volume inside B

n

1

is the Euclidean ball of radius

1

n

. (See

Figure 3.) There are 2

n

points of contact between this ball and the boundary of

B

n

1

: namely, the points of the form



±

1

n

,

±

1

n

, . . . ,

±

1

n



,

one in the middle of each facet of B

n

1

. The vertices,

(

±1, 0, 0, . . . , 0), . . . , (0, 0, . . . , 0, ±1),

are the points of B

n

1

furthest from the origin. We are looking for a rotation U B

n

1

whose facets chop off the spikes of B

n

1

(or vice versa). So we want to know that

the points of B

n

1

at distance about 1/

n from the origin are fairly typical, while

those at distance 1 are atypical.

For a unit vector θ

∈ S

n−1

, let r(θ) be the radius of B

n

1

in the direction θ,

r(θ) =

1

kθk

1

=



n

X

1

i

|



1

.

background image

AN ELEMENTARY INTRODUCTION TO MODERN CONVEX GEOMETRY

21

In the first lecture it was explained that the volume of B

n

1

can be written

v

n

Z

S

n−1

r(θ)

n

and that it is equal to 2

n

/n!. Hence

Z

S

n−1

r(θ)

n

=

2

n

n! v

n



2

n



n

.

Since the average of r(θ)

n

is at most 2/

n



n

, the value of r(θ) cannot often be

much more than 2/

n. This feature of B

n

1

is captured in the following definition

of Szarek.

Definition.

Let K be a convex body in

R

n

. The volume ratio of K is

vr(K) =



vol(K)

vol(

E)



1/n

,

where

E is the ellipsoid of maximal volume in K.

The preceding discussion shows that vr(B

n

1

)

2 for all n. Contrast this with the

cube in

R

n

, whose volume ratio is about

n/2. The only property of B

n

1

that

we shall use to prove Kaˇsin’s Theorem is that its volume ratio is at most 2. For
convenience, we scale everything up by a factor of

n and prove the following.

Theorem

4.2.

Let K be a symmetric convex body in

R

n

that contains the

Euclidean unit ball B

n

2

and for which



vol(K)

vol(B

n

2

)



1/n

= R.

Then there is an orthogonal transformation U of

R

n

for which

K

∩ UK ⊂ 8R

2

B

n

2

.

Proof.

It is convenient to work with the norm on

R

n

whose unit ball is K. Let

k · k denote this norm and | · | the standard Euclidean norm. Notice that, since
B

n

2

⊂ K, we have kxk ≤ |x| for all x ∈

R

n

.

The radius of the body K

∩ UK in a given direction is the minimum of the

radii of K and U K in that direction. So the norm corresponding to the body
K

∩ UK is the maximum of the norms corresponding to K and UK. We need

to find an orthogonal transformation U with the property that

max (

kUθk, kθk)

1

8R

2

for every θ

∈ S

n−1

. Since the maximum of two numbers is at least their average,

it will suffice to find U with

kUθk + kθk

2

1

8R

2

for all θ.

background image

22

KEITH BALL

For each x

R

n

write N (x) for the average

1

2

(

kUxk + kxk). One sees imme-

diately that N is a norm (that is, it satisfies the triangle inequality) and that
N (x)

≤ |x| for every x, since U preserves the Euclidean norm. We shall show in

a moment that there is a U for which

Z

S

n−1

1

N (θ)

2n

≤ R

2n

.

(4.1)

This says that N (θ) is large on average: we want to deduce that it is large
everywhere.

Let φ be a point of the sphere and write N (φ) = t, for 0 < t

1. The crucial

point will be that, if θ is close to φ, then N (θ) cannot be much more than t. To
be precise, if

|θ − φ| ≤ t then

N (θ)

≤ N(φ) + N(θ − φ) ≤ t + |θ − φ| ≤ 2t.

Hence, N (θ) is at most 2t for every θ in a spherical cap of radius t about φ.
From Lemma 2.3 we know that this spherical cap has measure at least

1

2



t

2



n−1



t

2



n

.

So 1/N (θ)

2n

is at least 1/(2t)

2n

on a set of measure at least (t/2)

n

. Therefore

Z

S

n−1

1

N (θ)

2n

1

(2t)

2n



t

2



n

=

1

2

3n

t

n

.

By (4.1), the integral is at most R

2n

, so t

1/(8R

2

). Thus our arbitrary point

φ satisfies

N (φ)

1

8R

2

.

It remains to find U satisfying (4.1). Now, for any θ, we have

N (θ)

2

=



kUθk + kθk

2



2

≥ kUθk kθk,

so it will suffice to find a U for which

Z

S

n−1

1

kUθk

n

kθk

n

≤ R

2n

.

(4.2)

The hypothesis on the volume of K can be written in terms of the norm as

Z

S

n−1

1

kθk

n

= R

n

.

The group of orthogonal transformations carries an invariant probability mea-
sure. This means that we can average a function over the group in a natural
way. In particular, if f is a function on the sphere and θ is some point on the

background image

AN ELEMENTARY INTRODUCTION TO MODERN CONVEX GEOMETRY

23

sphere, the average over orthogonal U of the value f (U θ) is just the average of
f on the sphere: averaging over U mimics averaging over the sphere:

ave

U

f (U θ) =

Z

S

n−1

f (φ) (φ).

Hence,

ave

U

Z

S

n−1

1

kUθk

n

.

kθk

n

(θ) =

Z

S

n−1



ave

U

1

kUθk

n



1

kθk

n

(θ)

=

Z

S

n−1

Z

S

n−1

1

kφk

n

(φ)



1

kθk

n

(θ)

=

Z

S

n−1

1

kθk

n

(θ)



2

= R

2n

.

Since the average over all U of the integral is at most R

2n

, there is at least one

U for which the integral is at most R

2n

. This is exactly inequality (4.2).



The choice of U in the preceding proof is a random one. The proof does not
in any way tell us how to find an explicit U for which the integral is small. In
the case of a general body K, this is hardly surprising, since we are assuming
nothing about how the volume of K is distributed. But, in view of the earlier
remarks about facets of B

n

1

chopping off spikes of U B

n

1

, it is tempting to think

that for the particular body B

n

1

we might be able to write down an appropriate

U explicitly. In two dimensions the best choice of U is obvious: we rotate the
diamond through 45

and after intersection we have a regular octagon as shown

in Figure 17.

The most natural way to try to generalise this to higher dimensions is to look

for a U such that each vertex of U B

n

1

is exactly aligned through the centre of a

facet of B

n

1

: that is, for each standard basis vector e

i

of

R

n

, U e

i

is a multiple of

one of the vectors (

±

1

n

, . . . ,

±

1

n

). (The multiple is

n since U e

i

has length 1.)

Thus we are looking for an n

× n orthogonal matrix U each of whose entries is

B

2

1

UB

2

1

B

2

1

∩ UB

2

1

Figure 17. The best choice for U in two dimensions is a 45

rotation.

background image

24

KEITH BALL

±1/

n. Such matrices, apart from the factor

n, are called Hadamard matrices.

In what dimensions do they exist? In dimensions 1 and 2 there are the obvious

(1)

and

1

2

1

2

1

2

1

2

!

.

It is not too hard to show that in larger dimensions a Hadamard matrix cannot
exist unless the dimension is a multiple of 4. It is an open problem to determine
whether they exist in all dimensions that are multiples of 4. They are known to
exist, for example, if the dimension is a power of 2: these examples are known
as the Walsh matrices.

In spite of this, it seems extremely unlikely that one might prove Kaˇsin’s

Theorem using Hadamard matrices. The Walsh matrices certainly do not give
anything smaller than n

1/4

; pretty miserable compared with n

1/2

. There

are some good reasons, related to Ramsey theory, for believing that one cannot
expect to find genuinely explicit matrices of any kind that would give the right
estimates.

Let’s return to the question with which we opened the lecture and see how

Theorem 4.1 yields almost spherical sections of octahedra. We shall show that,
for each n, the 2n-dimensional octahedron has an n-dimensional slice which
is within distance 32 of the (n-dimensional) Euclidean ball. By applying the
argument of the theorem to B

n

1

, we obtain an n

× n orthogonal matrix U such

that

kUxk

1

+

kxk

1

n

16

|x|

for every x

R

n

, where

k · k

1

denotes the `

1

norm. Now consider the map

T :

R

n

R

2n

with matrix

U

I



. For each x

R

n

, the norm of T x in `

2n

1

is

kT xk

1

=

kUxk

1

+

kxk

1

n

16

|x|.

On the other hand, the Euclidean norm of T x is

|T x| =

p

|Ux|

2

+

|x|

2

=

2

|x|.

So, if y belongs to the image T

R

n

, then, setting y = T x,

kyk

1

n

16

|x| =

n

16

2

|y| =

2n

32

|y|.

By the Cauchy–Schwarz inequality, we have

kyk

1

2n

|y|, so the slice of B

2n

1

by the subspace T

R

n

has distance at most 32 from B

n

2

, as we wished to show.

A good deal of work has been done on embedding of other subspaces of L

1

into `

1

-spaces of low dimension, and more generally subspaces of L

p

into low-

dimensional `

p

, for 1 < p < 2. The techniques used come from probability

theory: p-stable random variables, bootstrapping of probabilities and deviation
estimates. We shall be looking at applications of the latter in Lectures 8 and 9.

background image

AN ELEMENTARY INTRODUCTION TO MODERN CONVEX GEOMETRY

25

The major references are [Johnson and Schechtman 1982; Bourgain et al. 1989;
Talagrand 1990].

Volume ratios have been studied in considerable depth. They have been found

to be closely related to the so-called cotype-2 property of normed spaces: this re-
lationship is dealt with comprehensively in [Pisier 1989]. In particular, Bourgain
and Milman [1987] showed that a bound for the cotype-2 constant of a space
implies a bound for the volume ratio of its unit ball. This demonstrated, among
other things, that there is a uniform bound for the volume ratios of slices of
octahedra of all dimensions. A sharp version of this result was proved in [Ball
1991]: namely, that for each n, B

n

1

has largest volume ratio among the balls of

n-dimensional subspaces of L

1

. The proof uses techniques that will be discussed

in Lecture 6.

This is a good point to mention a result of Milman [1985] that looks super-

ficially like the results of this lecture but lies a good deal deeper. We remarked
that while we can almost get a sphere by intersecting two copies of B

n

1

, this is

very far from possible with two cubes. Conversely, we can get an almost spher-
ical convex hull of two cubes but not of two copies of B

n

1

. The QS-Theorem

(an abbreviation for “quotient of a subspace”) states that if we combine the two
operations, intersection and convex hull, we can get a sphere no matter what
body we start with.

Theorem 4.3 (QS-Theorem).

There is a constant M (independent of ev-

erything) such that , for any symmetric convex body K of any dimension, there
are linear maps Q and S and an ellipsoid

E with the following property: if

˜

K = conv(K

∪ QK), then

E ⊂ ˜

K

∩ S ˜

K

⊂ ME.

Lecture 5. The Brunn–Minkowski Inequality

and Its Extensions

In this lecture we shall introduce one of the most fundamental principles in

convex geometry: the Brunn–Minkowski inequality. In order to motivate it, let’s
begin with a simple observation concerning convex sets in the plane. Let K

R

2

be such a set and consider its slices by a family of parallel lines, for example those
parallel to the y-axis. If the line x = r meets K, call the length of the slice v(r).

The graph of v is obtained by shaking K down onto the x-axis like a deck of

cards (of different lengths). This is shown in Figure 18. It is easy to see that the
function v is concave on its support. Towards the end of the last century, Brunn
investigated what happens if we try something similar in higher dimensions.

Figure 19 shows an example in three dimensions. The central, hexagonal, slice

has larger volume than the triangular slices at the ends: each triangular slice
can be decomposed into four smaller triangles, while the hexagon is a union of
six such triangles. So our first guess might be that the slice area is a concave

background image

26

KEITH BALL

v(x)

K

Figure 18. Shaking down a convex body.

function, just as slice length was concave for sets in the plane. That this is not
always so can be seen by considering slices of a cone, parallel to its base: see
Figure 20.
Since the area of a slice varies as the square of its distance from the cone’s
vertex, the area function obtained looks like a piece of the curve y = x

2

, which

is certainly not concave. However, it is reasonable to guess that the cone is an
extremal example, since it is “only just” a convex body: its curved surface is
“made up of straight lines”. For this body, the square root of the slice function
just manages to be concave on its support (since its graph is a line segment). So
our second guess might be that for a convex body in

R

3

, a slice-area function has

a square-root that is concave on its support. This was proved by Brunn using
an elegant symmetrisation method. His argument works just as well in higher
dimensions to give the following result for the (n

1)-dimensional volumes of

slices of a body in

R

n

.

Figure 19. A polyhedron in three dimensions. The faces at the right and left
are parallel.

background image

AN ELEMENTARY INTRODUCTION TO MODERN CONVEX GEOMETRY

27

x

Figure 20. The area of a cone’s section increases with x

2

.

Theorem 5.1 (Brunn).

Let K be a convex body in

R

n

, let u be a unit vector

in

R

n

, and for each r let H

r

be the hyperplane

{x ∈

R

n

:

hx, ui = r} .

Then the function

r

7→ vol (K ∩ H

r

)

1/(n−1)

is concave on its support .

One consequence of this is that if K is centrally symmetric, the largest slice
perpendicular to a given direction is the central slice (since an even concave
function is largest at 0). This is the situation in Figure 19.

Brunn’s Theorem was turned from an elegant curiosity into a powerful tool by

Minkowski. His reformulation works in the following way. Consider three parallel
slices of a convex body in

R

n

at positions r, s and t, where s = (1

− λ)r + λt for

some λ

(0, 1). This is shown in Figure 21.

Call the slices A

r

, A

s

, and A

t

and think of them as subsets of

R

n−1

. If x

∈ A

r

and y

∈ A

t

, the point (1

− λ)x + λy belongs to A

s

: to see this, join the points

(r, x) and (t, y) in

R

n

and observe that the resulting line segment crosses A

s

at

(s, (1

− λ)x + λy). So A

s

includes a new set

(1

− λ)A

r

+ λA

t

:=

{(1 − λ)x + λy : x ∈ A

r

, y

∈ A

t

} .

A

r

A

s

A

t

(t, y)

(s, x)

Figure 21. The section A

s

contains the weighted average of A

r

and A

t

.

background image

28

KEITH BALL

(This way of using the addition in

R

n

to define an addition of sets is called

Minkowski addition.) Brunn’s Theorem says that the volumes of the three sets
A

r

, A

s

, and A

t

in

R

n−1

satisfy

vol (A

s

)

1/(n−1)

(1 − λ) vol (A

r

)

1/(n−1)

+ λ vol (A

t

)

1/(n−1)

.

The Brunn–Minkowski inequality makes explicit the fact that all we really know
about A

s

is that it includes the Minkowski combination of A

r

and A

t

. Since we

have now eliminated the role of the ambient space

R

n

, it is natural to rewrite

the inequality with n in place of n

1.

Theorem 5.2 (Brunn–Minkowski inequality).

If A and B are nonempty

compact subsets of

R

n

then

vol ((1

− λ)A + λB)

1/n

(1 − λ) vol (A)

1/n

+ λ vol (B)

1/n

.

(The hypothesis that A and B be nonempty corresponds in Brunn’s Theorem
to the restriction of a function to its support.) It should be remarked that the
inequality is stated for general compact sets, whereas the early proofs gave the
result only for convex sets. The first complete proof for the general case seems
to be in [Lıusternik 1935].

To get a feel for the advantages of Minkowski’s formulation, let’s see how it

implies the classical isoperimetric inequality in

R

n

.

Theorem 5.3 (Isoperimetric inequality).

Among bodies of a given volume,

Euclidean balls have least surface area.

Proof.

Let C be a compact set in

R

n

whose volume is equal to that of B

n

2

, the

Euclidean ball of radius 1. The surface “area” of C can be written

vol(∂C) = lim

ε→0

vol (C + εB

n

2

)

vol (C)

ε

,

as shown in Figure 22. By the Brunn–Minkowski inequality,

vol (C + εB

n

2

)

1/n

vol (C)

1/n

+ ε vol (B

n

2

)

1/n

.

K

K + εB

n

2

Figure 22. Expressing the area as a limit of volume increments.

background image

AN ELEMENTARY INTRODUCTION TO MODERN CONVEX GEOMETRY

29

Hence

vol (C + εB

n

2

)



vol (C)

1/n

+ ε vol (B

n

2

)

1/n



n

vol (C) + vol (C)

(n−1)/n

vol (B

n

2

)

1/n

.

So

vol(∂C)

≥ n vol(C)

(n−1)/n

vol(B

n

2

)

1/n

.

Since C and B

n

2

have the same volume, this shows that vol(∂C)

≥ n vol(B

n

2

),

and the latter equals vol(∂B

n

2

), as we saw in Lecture 1.



This relationship between the Brunn–Minkowski inequality and the isoperimetric
inequality will be explored in a more general context in Lecture 8.

The Brunn–Minkowski inequality has an alternative version that is formally

weaker. The AM/GM inequality shows that, for λ in (0, 1),

(1

− λ) vol(A)

1/n

+ λ vol(B)

1/n

vol(A)

(1−λ)/n

vol(B)

λ/n

.

So the Brunn–Minkowski inequality implies that, for compact sets A and B and
λ

(0, 1),

vol((1

− λ)A + λB) vol(A)

1−λ

vol(B)

λ

.

(5.1)

Although this multiplicative Brunn–Minkowski inequality is weaker than the
Brunn–Minkowski inequality for particular A, B, and λ, if one knows (5.1) for
all A, B, and λ one can easily deduce the Brunn–Minkowski inequality for all
A, B, and λ. This deduction will be left for the reader.

Inequality (5.1) has certain advantages over the Brunn–Minkowski inequality.

(i) We no longer need to stipulate that A and B be nonempty, which makes the

inequality easier to use.

(ii) The dimension n has disappeared.
(iii) As we shall see, the multiplicative inequality lends itself to a particularly

simple proof because it has a generalisation from sets to functions.

Before we describe the functional Brunn–Minkowski inequality let’s just remark
that the multiplicative Brunn–Minkowski inequality can be reinterpreted back in
the setting of Brunn’s Theorem: if r

7→ v(r) is a function obtained by scanning

a convex body with parallel hyperplanes, then log v is a concave function (with
the usual convention regarding

−∞).

In order to move toward a functional generalisation of the multiplicative

Brunn–Minkowski inequality let’s reinterpret inequality (5.1) in terms of the
characteristic functions of the sets involved. Let f , g, and m denote the char-
acteristic functions of A, B, and (1

− λ)A + λB respectively; so, for example,

f (x) = 1 if x

∈ A and 0 otherwise. The volumes of A, B, and (1 − λ)A + λB

are the integrals

R

R

n

f ,

R

R

n

g, and

R

R

n

m. The Brunn–Minkowski inequality says

that

Z

m

Z

f



1−λ

Z

g



λ

.

background image

30

KEITH BALL

But what is the relationship between f , g, and m that guarantees its truth? If
f (x) = 1 and g(y) = 1 then x

∈ A and y ∈ B, so

(1

− λ)x + λy ∈ (1 − λ)A + λB,

and hence m ((1

− λ)x + λy) = 1. This certainly ensures that

m ((1

− λ)x + λy) ≥ f(x)

1−λ

g(y)

λ

for any x and y in

R

n

.

This inequality for the three functions at least has a homogeneity that matches
the desired inequality for the integrals.

In a series of papers, Pr´

ekopa and

Leindler proved that this homogeneity is enough.

Theorem 5.4 (The Pr´

ekopa–Leindler inequality).

If f , g and m are

nonnegative measurable functions on

R

n

, λ

(0, 1) and for all x and y in

R

n

,

m ((1

− λ)x + λy) ≥ f(x)

1−λ

g(y)

λ

(5.2)

then

Z

m

Z

f



1−λ

Z

g



λ

.

It is perhaps helpful to notice that the Pr´

ekopa–Leindler inequality looks like

older’s inequality, backwards. If f and g were given and we set

m(z) = f (z)

1−λ

g(z)

λ

(for each z), then H¨

older’s inequality says that

Z

m

Z

f



1−λ

Z

g



λ

.

(H¨

older’s inequality is often written with 1/p instead of 1

− λ, 1/q instead of

λ, and f , g replaced by F

p

, G

q

.) The difference between Pr´

ekopa–Leindler and

older is that, in the former, the value m(z) may be much larger since it is a

supremum over many pairs (x, y) satisfying z = (1

− λ)x + λy rather than just

the pair (z, z).

Though it generalises the Brunn–Minkowski inequality, the Pr´

ekopa–Leindler

inequality is a good deal simpler to prove, once properly formulated. The argu-
ment we shall use seems to have appeared first in [Brascamp and Lieb 1976b].
The crucial point is that the passage from sets to functions allows us to prove
the inequality by induction on the dimension, using only the one-dimensional
case. We pay the small price of having to do a bit extra for this case.

Proof of the Pr´

ekopa–Leindler inequality.

We start by checking the

one-dimensional Brunn–Minkowski inequality. Suppose A and B are nonempty
measurable subsets of the line. Using

| · | to denote length, we want to show that

|(1 − λ)A + λB| ≥ (1 − λ)|A| + λ|B|.

background image

AN ELEMENTARY INTRODUCTION TO MODERN CONVEX GEOMETRY

31

We may assume that A and B are compact and we may shift them so that
the right-hand end of A and the left-hand end of B are both at 0. The set
(1

− λ)A + λB now includes the essentially disjoint sets (1 − λ)A and λB, so its

length is at least the sum of the lengths of these sets.

Now suppose we have nonnegative integrable functions f , g, and m on the

line, satisfying condition (5.2). We may assume that f and g are bounded. Since
the inequality to be proved has the same homogeneity as the hypothesis (5.2),
we may also assume that f and g are normalised so that sup f = sup g = 1.
By Fubini’s Theorem, we can write the integrals of f and g as integrals of the
lengths of their level sets:

Z

f (x) dx =

Z

1

0

|(f ≥ t)| dt,

and similarly for g. If f (x)

≥ t and g(y) ≥ t then m ((1 − λ)x + λy) ≥ t. So we

have the inclusion

(m

≥ t) (1 − λ)(f ≥ t) + λ(g ≥ t).

For 0

≤ t < 1 the sets on the right are nonempty so the one-dimensional Brunn–

Minkowski inequality shows that

|(m ≥ t)| ≥ (1 − λ) |(f ≥ t)| + λ |(g ≥ t)| .

Integrating this inequality from 0 to 1 we get

Z

m

(1 − λ)

Z

f + λ

Z

g,

and the latter expression is at least

Z

f



1−λ

Z

g



λ

by the AM/GM inequality. This does the one-dimensional case.

The induction that takes us into higher dimensions is quite straightforward, so

we shall just sketch the argument for sets in

R

n

, rather than functions. Suppose

A and B are two such sets and, for convenience, write

C = (1

− λ)A + λB.

Choose a unit vector u and, as before, let H

r

be the hyperplane

{x ∈

R

n

:

hx, ui = r}

perpendicular to u at “position” r. Let A

r

denote the slice A

∩ H

r

and similarly

for B and C, and regard these as subsets of

R

n−1

. If r and t are real numbers,

and if s = (1

−λ)r+λt, the slice C

s

includes (1

−λ)A

r

+λB

t

. (This is reminiscent

background image

32

KEITH BALL

of the earlier argument relating Brunn’s Theorem to Minkowski’s reformulation.)
By the inductive hypothesis in

R

n−1

,

vol(C

s

)

vol(A

r

)

1−λ

. vol(B

t

)

λ

.

Let f , g, and m be the functions on the line, given by

f (x) = vol(A

x

),

g(x) = vol(B

x

),

m(x) = vol(C

x

).

Then, for r, s, and t as above,

m(s)

≥ f(r)

1−λ

g(t)

λ

.

By the one-dimensional Pr´

ekopa–Leindler inequality,

Z

m

Z

f



1−λ

Z

g



λ

.

But this is exactly the statement vol(C)

vol(A)

1−λ

vol(B)

λ

, so the inductive

step is complete.



The proof illustrates clearly why the Pr´

ekopa–Leindler inequality makes things

go smoothly. Although we only carried out the induction for sets, we required
the one-dimensional result for the functions we get by scanning sets in

R

n

.

To close this lecture we remark that the Brunn–Minkowski inequality has nu-

merous extensions and variations, not only in convex geometry, but in combina-
torics and information theory as well. One of the most surprising and delightful
is a theorem of Busemann [1949].

Theorem 5.5 (Busemann).

Let K be a symmetric convex body in

R

n

, and

for each unit vector u let r(u) be the volume of the slice of K by the subspace
orthogonal to u
. Then the body whose radius in each direction u is r(u) is itself
convex
.

The Brunn–Minkowski inequality is the starting point for a highly developed
classical theory of convex geometry. We shall barely touch upon the theory in
these notes. A comprehensive reference is the recent book [Schneider 1993].

Lecture 6. Convolutions and Volume Ratios:

The Reverse Isoperimetric Problem

In the last lecture we saw how to deduce the classical isoperimetric inequality

in

R

n

from the Brunn–Minkowski inequality. In this lecture we will answer the

reverse question. This has to be phrased a bit carefully, since there is no upper
limit to the surface area of a body of given volume, even if we restrict attention
to convex bodies. (Consider a very thin pancake.) For this reason it is natural to
consider affine equivalence classes of convex bodies, and the question becomes:
given a convex body, how small can we make its surface area by applying an

background image

AN ELEMENTARY INTRODUCTION TO MODERN CONVEX GEOMETRY

33

affine (or linear) transformation that preserves volume? The answer is provided
by the following theorem from [Ball 1991].

Theorem 6.1.

Let K be a convex body and T a regular solid simplex in

R

n

.

Then there is an affine image of K whose volume is the same as that of T and
whose surface area is no larger than that of T
.

Thus, modulo affine transformations, simplices have the largest surface area
among convex bodies of a given volume. If K is assumed to be centrally sym-
metric then the estimate can be strengthened: the cube is extremal among sym-
metric bodies. A detailed proof of Theorem 6.1 would be too long for these
notes. We shall instead describe how the symmetric case is proved, since this is
considerably easier but illustrates the most important ideas.

Theorem 6.1 and the symmetric analogue are both deduced from volume-ratio

estimates. In the latter case the statement is that among symmetric convex
bodies, the cube has largest volume ratio. Let’s see why this solves the reverse
isoperimetric problem. If Q is any cube, the surface area and volume of Q are
related by

vol(∂Q) = 2n vol(Q)

(n−1)/n

.

We wish to show that any other convex body K has an affine image ˜

K for which

vol(˜

K)

2n vol( ˜

K)

(n−1)/n

.

Choose ˜

K so that its maximal volume ellipsoid is B

n

2

, the Euclidean ball of

radius 1. The volume of ˜

K is then at most 2

n

, since this is the volume of the

cube whose maximal ellipsoid is B

n

2

. As in the previous lecture,

vol(˜

K) = lim

ε→0

vol( ˜

K + εB

n

2

)

vol( ˜

K)

ε

.

Since ˜

K

⊃ B

n

2

, the second expression is at most

lim

ε→0

vol( ˜

K + ε ˜

K)

vol( ˜

K)

ε

= vol( ˜

K) lim

ε→0

(1 + ε)

n

1

ε

= n vol( ˜

K) = n vol( ˜

K)

1/n

vol( ˜

K)

(n−1)/n

2n vol( ˜

K)

(n−1)/n

,

which is exactly what we wanted.

The rest of this lecture will thus be devoted to explaining the proof of the

volume-ratio estimate:

Theorem 6.2.

Among symmetric convex bodies the cube has largest volume

ratio.

As one might expect, the proof of Theorem 6.2 makes use of John’s Theorem
from Lecture 3. The problem is to show that, if K is a convex body whose
maximal ellipsoid is B

n

2

, then vol(K)

2

n

. As we saw, it is a consequence of

background image

34

KEITH BALL

John’s theorem that if B

n

2

is the maximal ellipsoid in K, there is a sequence (u

i

)

of unit vectors and a sequence (c

i

) of positive numbers for which

X

c

i

u

i

⊗ u

i

= I

n

and for which

K

⊂ C := {x : |hx, u

i

i| ≤ 1 for 1 ≤ i ≤ m} .

We shall show that this C has volume at most 2

n

. The principal tool will be a

sharp inequality for norms of generalised convolutions. Before stating this let’s
explain some standard terms from harmonic analysis.

If f and g :

R

R

are bounded, integrable functions, we define the convolu-

tion f

∗ g of f and g by

f

∗ g(x) =

Z

R

f (y)g(x

− y) dy.

Convolutions crop up in many areas of mathematics and physics, and a good
deal is known about how they behave. One of the most fundamental inequalities
for convolutions is Young’s inequality: If f

∈ L

p

, g

∈ L

q

, and

1

p

+

1

q

= 1 +

1

s

,

then

kf ∗ gk

s

≤ kfk

p

kgk

q

.

(Here

k · k

p

means the L

p

norm on

R

, and so on.) Once we have Young’s in-

equality, we can give a meaning to convolutions of functions that are not both
integrable and bounded, provided that they lie in the correct L

p

spaces. Young’s

inequality holds for convolution on any locally compact group, for example the
circle. On compact groups it is sharp: there is equality for constant functions.
But on

R

, where constant functions are not integrable, the inequality can be

improved (for most values of p and q). It was shown by Beckner [1975] and
Brascamp and Lieb [1976a] that the correct constant in Young’s inequality is
attained if f and g are appropriate Gaussian densities: that is, for some positive
a and b, f (t) = e

−at

2

and g(t) = e

−bt

2

. (The appropriate choices of a and b and

the value of the best constant for each p and q will not be stated here. Later we
shall see that they can be avoided.)

How are convolutions related to convex bodies? To answer this question we

need to rewrite Young’s inequality slightly. If 1/r +1/s = 1, the L

s

norm

kf ∗gk

s

can be realised as

Z

R

(f

∗ g)(x)h(x)

for some function h with

khk

r

= 1. So the inequality says that, if 1/p+1/q+1/r =

2, then

Z Z

f (y)g(x

− y)h(x) dy dx ≤ kfk

p

kgk

q

khk

r

.

background image

AN ELEMENTARY INTRODUCTION TO MODERN CONVEX GEOMETRY

35

We may rewrite the inequality again with h(

−x) in place of h(x), since this

doesn’t affect

khk

r

:

Z Z

f (y)g(x

− y)h(−x) dy dx ≤ kfk

p

kgk

q

khk

r

.

(6.1)

This can be written in a more symmetric form via the map from

R

2

into

R

3

that

takes (x, y) to (y, x

−y, −x) =: (u, v, w). The range of this map is the subspace

H =

{(u, v, w) : u + v + w = 0} .

Apart from a factor coming from the Jacobian of this map, the integral can be
written

Z

H

f (u)g(v)h(w),

where the integral is with respect to two-dimensional measure on the subspace
H. So Young’s inequality and its sharp forms estimate the integral of a product
function on

R

3

over a subspace. What is the simplest product function? If f , g,

and h are each the characteristic function of the interval [

1, 1], the function F

given by

F (u, v, w) = f (u)g(v)h(w)

is the characteristic function of the cube [

1, 1]

3

R

3

. The integral of F over

a subspace of

R

3

is thus the area of a slice of the cube: the area of a certain

convex body. So there is some hope that we might use a convolution inequality
to estimate volumes.

Brascamp and Lieb proved rather more than the sharp form of Young’s in-

equality stated earlier. They considered not just two-dimensional subspaces of

R

3

but n-dimensional subspaces of

R

m

. It will be more convenient to state their

result using expressions analogous to those in (6.1) rather than using integrals
over subspaces. Notice that the integral

Z Z

f (y)g(x

− y)h(−x) dy dx

can be written

Z

R

2

f

hx, v

1

i



g

hx, v

2

i



h

hx, v

3

i



dx,

where v

1

= (0, 1), v

2

= (1,

1) and v

3

= (

1, 0) are vectors in

R

2

. The theorem

of Brascamp and Lieb is the following.

Theorem 6.3.

If (v

i

)

m

1

are vectors in

R

n

and (p

i

)

m

1

are positive numbers satis-

fying

m

X

1

1

p

i

= n,

and if (f

i

)

m

1

are nonnegative measurable functions on the line, then

R

R

n

Q

m

1

f

i

(

hx, v

i

i)

Q

m

1

kf

i

k

p

i

background image

36

KEITH BALL

is “maximised” when the (f

i

) are appropriate Gaussian densities : f

i

(t) = e

−a

i

t

2

,

where the a

i

depend upon m, n, the p

i

, and the v

i

.

The word maximised is in quotation marks since there are degenerate cases for
which the maximum is not attained. The value of the maximum is not easily
computed since the a

i

are the solutions of nonlinear equations in the p

i

and

v

i

. This apparently unpleasant problem evaporates in the context of convex

geometry: the inequality has a normalised form, introduced in [Ball 1990], which
fits perfectly with John’s Theorem.

Theorem 6.4.

If (u

i

)

m

1

are unit vectors in

R

n

and (c

i

)

m

1

are positive numbers

for which

m

X

1

c

i

u

i

⊗ u

i

= I

n

,

and if (f

i

)

m

1

are nonnegative measurable functions, then

Z

R

n

Y

f

i

(

hx, u

i

i)

c

i

Y Z

f

i



c

i

.

In this reformulation of the theorem, the c

i

play the role of 1/p

i

: the Fritz John

condition ensures that

P

c

i

= n as required, and miraculously guarantees that

the correct constant in the inequality is 1 (as written). The functions f

i

have

been replaced by f

c

i

i

, since this ensures that equality occurs if the f

i

are identical

Gaussian densities. It may be helpful to see why this is so. If f

i

(t) = e

−t

2

for all

i, then

Y

f

i

(

hx, u

i

i)

c

i

= exp



X

c

i

hx, u

i

i

2



= e

−|x|

2

=

n

Y

1

e

−x

2

i

,

so the integral is

Z

e

−t

2



n

=

Y Z

e

−t

2



c

i

=

Y Z

f

i



c

i

.

Armed with Theorem 6.4, let’s now prove Theorem 6.2.

Proof of the volume-ratio estimate.

Recall that our aim is to show that,

for u

i

and c

i

as usual, the body

C =

{x : |hx, u

i

i| ≤ 1 for 1 ≤ i ≤ m}

has volume at most 2

n

. For each i let f

i

be the characteristic function of the

interval [

1, 1] in

R

. Then the function

x

7→

Y

f

i

hx, u

i

i



c

i

background image

AN ELEMENTARY INTRODUCTION TO MODERN CONVEX GEOMETRY

37

is exactly the characteristic function of C. Integrating and applying Theorem
6.4 we have

vol(C)

Y Z

f

i



c

i

=

Y

2

c

i

= 2

n

.



The theorems of Brascamp and Lieb and Beckner have been greatly extended
over the last twenty years. The main result in Beckner’s paper solved the old
problem of determining the norm of the Fourier transform between L

p

spaces.

There are many classical inequalities in harmonic analysis for which the best
constants are now known. The paper [Lieb 1990] contains some of the most
up-to-date discoveries and gives a survey of the history of these developments.

The methods described here have many other applications to convex geometry.

There is also a reverse form of the Brascamp–Lieb inequality appropriate for
analysing, for example, the ratio of the volume of a body to that of the minimal
ellipsoid containing it.

Lecture 7. The Central Limit Theorem

and Large Deviation Inequalities

The material in this short lecture is not really convex geometry, but is intended

to provide a context for what follows. For the sake of readers who may not be
familiar with probability theory, we also include a few words about independent
random variables.

To begin with, a probability measure µ on a set Ω is just a measure of total

mass µ(Ω) = 1. Real-valued functions on Ω are called random variables and
the integral of such a function X : Ω

R

, its mean, is written EX and called

the expectation of X. The variance of X is E(X

− EX)

2

. It is customary to

suppress the reference to Ω when writing the measures of sets defined by random
variables. Thus

µ(

{ω ∈ Ω : X(ω) < 1})

is written µ(X < 1): the probability that X is less than 1.

Two crucial, and closely related, ideas distinguish probability theory from

general measure theory. The first is independence. Two random variables X
and Y are said to be independent if, for any functions f and g,

Ef (X)g(Y ) = Ef (X) Eg(Y ).

Independence can always be viewed in a canonical way. Let (Ω, µ) be a product
space (Ω

1

×

2

, µ

1

⊗ µ

2

), where µ

1

and µ

2

are probabilities. Suppose X and

Y are random variables on Ω for which the value X(ω

1

, ω

2

) depends only upon

ω

1

while Y (ω

1

, ω

2

) depends only upon ω

2

. Then any integral (that converges

appropriately)

Ef (X)g(Y ) =

Z

f (X(s)) g(Y (t))

1

⊗ µ

2

(s, t)

background image

38

KEITH BALL

s

0

2

1

Figure 23. Independence and product spaces

can be written as the product of integrals

Z

f (X(s))

1

(s)

Z

g(Y (t))

2

(t) = Ef (X)Eg(Y )

by Fubini’s Theorem. Putting it another way, on each line

{(s

0

, t) : t

2

},

X is fixed, while Y exhibits its full range of behaviour in the correct proportions.
This is illustrated in Figure 23.

In a similar way, a sequence X

1

, X

2

, . . . , X

n

of independent random variables

arises if each variable is defined on the product space Ω

1

×

2

× . . . ×

n

and X

i

depends only upon the i-th coordinate.

The second crucial idea, which we will not discuss in any depth, is the use

of many different σ-fields on the same space. The simplest example has already
been touched upon. The product space Ω

1

×

2

carries two σ-fields, much smaller

than the product field, which it inherits from Ω

1

and Ω

2

respectively. If

F

1

and

F

2

are the σ-fields on Ω

1

and Ω

2

, the sets of the form A

×

2

1

×

2

for

A

∈ F

1

form a σ-field on Ω

1

×

2

; let’s call it ˜

F

1

. Similarly,

˜

F

2

=

{

1

× B : B ∈ F

2

}.

“Typical” members of these σ-fields are shown in Figure 24.

˜

F

1

˜

F

2

Figure 24. Members of the “small” σ-fields ˜

F

1

and ˜

F

2

on Ω

1

×

2

.

background image

AN ELEMENTARY INTRODUCTION TO MODERN CONVEX GEOMETRY

39

One of the most beautiful and significant principles in mathematics is the cen-

tral limit theorem: any random quantity that arises as the sum of many small
independent contributions is distributed very much like a Gaussian random vari-
able. The most familiar example is coin tossing. We use a coin whose decoration
is a bit austere: it has +1 on one side and

1 on the other. Let ε

1

, ε

2

, . . . , ε

n

be the outcomes of n independent tosses. Thus the ε

i

are independent random

variables, each of which takes the values +1 and

1 each with probability

1

2

.

(Such random variables are said to have a Bernoulli distribution.) Then the
normalised sum

S

n

=

1

n

n

X

1

ε

i

belongs to an interval I of the real line with probability very close to

1

2π

Z

I

e

−t

2

/2

dt.

The normalisation 1/

n, ensures that the variance of S

n

is 1: so there is some

hope that the S

n

will all be similarly distributed.

The standard proof of the central limit theorem shows that much more is true.

Any sum of the form

n

X

1

a

i

ε

i

with real coefficients a

i

will have a roughly Gaussian distribution as long as each

a

i

is fairly small compared with

P

a

i

2

. Some such smallness condition is clearly

needed since if

a

1

= 1

and

a

2

= a

3

=

· · · = a

n

= 0,

the sum is just ε

1

, which is not much like a Gaussian. However, in many in-

stances, what one really wants is not that the sum is distributed like a Gaussian,
but merely that the sum cannot be large (or far from average) much more often
than an appropriate Gaussian variable. The example above clearly satisfies such
a condition: ε

1

never deviates from its mean, 0, by more than 1.

The following inequality provides a deviation estimate for any sequence of

coefficients. In keeping with the custom among functional analysts, I shall refer
to the inequality as Bernstein’s inequality. (It is not related to the Bernstein
inequality for polynomials on the circle.) However, probabilists know the result
as Hoeffding’s inequality, and the earliest reference known to me is [Hoeffding
1963]. A stronger and more general result goes by the name of the Azuma–
Hoeffding inequality; see [Williams 1991], for example.

Theorem 7.1 (Bernstein’s inequality).

If ε

1

, ε

2

, . . . , ε

n

are independent

Bernoulli random variables and if a

1

, a

2

, . . . , a

n

satisfy

P

a

i

2

= 1, then for each

positive t we have

Prob



n

X

i=1

a

i

ε

i

> t



2e

−t

2

/2

.

background image

40

KEITH BALL

This estimate compares well with the probability of finding a standard Gaussian
outside the interval [

−t, t],

2

2π

Z

t

e

−s

2

/2

ds.

The method by which Bernstein’s inequality is proved has become an industry
standard.

Proof.

We start by showing that, for each real λ,

Ee

λ

P

a

i

ε

i

≤ e

λ

2

/2

.

(7.1)

The idea will then be that

P

a

i

ε

i

cannot be large too often, since, whenever it

is large, its exponential is enormous.

To prove (7.1), we write

Ee

λ

P

a

i

ε

i

= E

n

Y

1

e

λa

i

ε

i

and use independence to deduce that this equals

n

Y

1

Ee

λa

i

ε

i

.

For each i the expectation is

Ee

λa

i

ε

i

=

e

λa

i

+ e

−λa

i

2

= cosh λa

i

.

Now, cosh x

≤ e

x

2

/2

for any real x, so, for each i,

Ee

λa

i

ε

i

≤ e

λ

2

a

2

i

/2

.

Hence

Ee

λ

P

a

i

ε

i

n

Y

1

e

λ

2

a

2

i

/2

= e

λ

2

/2

,

since

P

a

2

i

= 1.

To pass from (7.1) to a probability estimate, we use the inequality variously

known as Markov’s or Chebyshev’s inequality: if X is a nonnegative random
variable and R is positive, then

R Prob(X

≥ R) ≤ EX

(because the integral includes a bit where a function whose value is at least R is
integrated over a set of measure Prob(X

≥ R)).

Suppose t

0. Whenever

P

a

i

ε

i

≥ t, we will have e

t

P

a

i

ε

i

≥ e

t

2

. Hence

e

t

2

Prob

X

a

i

ε

i

≥ t



≤ Ee

t

P

a

i

ε

i

≤ e

t

2

/2

background image

AN ELEMENTARY INTRODUCTION TO MODERN CONVEX GEOMETRY

41

by (7.1). So

Prob

X

a

i

ε

i

≥ t



≤ e

−t

2

/2

,

and in a similar way we get

Prob

X

a

i

ε

i

≤ −t



≤ e

−t

2

/2

.

Putting these inequalities together we get

Prob



X

a

i

ε

i

≥ t



2e

−t

2

/2

.



In the next lecture we shall see that deviation estimates that look like Bernstein’s
inequality hold for a wide class of functions in several geometric settings. For
the moment let’s just remark that an estimate similar to Bernstein’s inequality,

Prob



X

a

i

X

i

≥ t



2e

6t

2

,

(7.2)

holds for

P

a

2

i

= 1, if the

±1 valued random variables ε

i

are replaced by in-

dependent random variables X

i

each of which is uniformly distributed on the

interval



1

2

,

1

2



. This already has a more geometric flavour, since for these (X

i

)

the vector (X

1

, X

2

, . . . , X

n

) is distributed according to Lebesgue measure on

the cube



1

2

,

1

2



n

R

n

. If

P

a

2

i

= 1, then

P

a

i

X

i

is the distance of the point

(X

1

, X

2

, . . . , X

n

) from the subspace of

R

n

orthogonal to (a

1

, a

2

, . . . , a

n

). So (7.2)

says that most of the mass of the cube lies close to any subspace of

R

n

, which is

reminiscent of the situation for the Euclidean ball described in Lecture 1.

Lecture 8. Concentration of Measure in Geometry

The aim of this lecture is to describe geometric analogues of Bernstein’s de-

viation inequality. These geometric deviation estimates are closely related to
isoperimetric inequalities. The phenomenon of which they form a part was in-
troduced into the field by V. Milman: its development, especially by Milman
himself, led to a new, probabilistic, understanding of the structure of convex
bodies in high dimensions. The phenomenon was aptly named the concentration
of measure.

We explained in Lecture 5 how the Brunn–Minkowski inequality implies the

classical isoperimetric inequality in

R

n

: among bodies of a given volume, the

Euclidean balls have least surface area. There are many other situations where
isoperimetric inequalities are known; two of them will be described below. First
let’s recall that the argument from the Brunn–Minkowski inequality shows more
than the isoperimetric inequality.

Let A be a compact subset of

R

n

. For each point x of

R

n

, let d(x, A) be the

distance from x to A:

d(x, A) = min

{|x − y| : y ∈ A} .

background image

42

KEITH BALL

A

A

ε

ε

Figure 25. An ε-neighbourhood.

For each positive ε, the Minkowski sum A + εB

n

2

is exactly the set of points

whose distance from A is at most ε. Let’s denote such an ε-neighbourhood A

ε

;

see Figure 25.

The Brunn–Minkowski inequality shows that, if B is an Euclidean ball of the

same volume as A, we have

vol(A

ε

)

vol(B

ε

)

for any ε > 0.

This formulation of the isoperimetric inequality makes much clearer the fact that
it relates the measure and the metric on

R

n

. If we blow up a set in

R

n

using the

metric, we increase the measure by at least as much as we would for a ball.

This idea of comparing the volumes of a set and its neighbourhoods makes

sense in any space that has both a measure and a metric, regardless of whether
there is an analogue of Minkowski addition. For any metric space (Ω, d) equipped
with a Borel measure µ, and any positive α and ε, it makes sense to ask: For
which sets A of measure α do the blow-ups A

ε

have smallest measure? This

general isoperimetric problem has been solved in a variety of different situations.
We shall consider two closely related geometric examples. In each case the mea-
sure µ will be a probability measure: as we shall see, in this case, isoperimetric
inequalities may have totally unexpected consequences.

In the first example, Ω will be the sphere S

n−1

in

R

n

, equipped with either

the geodesic distance or, more simply, the Euclidean distance inherited from

R

n

as shown in Figure 26. (This is also called the chordal metric; it was used

in Lecture 2 when we discussed spherical caps of given radii.) The measure
will be σ = σ

n−1

, the rotation-invariant probability on S

n−1

. The solutions of

the isoperimetric problem on the sphere are known exactly: they are spherical
caps (Figure 26, right) or, equivalently, they are balls in the metric on S

n−1

.

Thus, if a subset A of the sphere has the same measure as a cap of radius r, its
neighbourhood A

ε

has measure at least that of a cap of radius r + ε.

This statement is a good deal more difficult to prove than the classical isoperi-

metric inequality on

R

n

: it was discovered by P. L´

evy, quite some time after the

isoperimetric inequality in

R

n

. At first sight, the statement looks innocuous

background image

AN ELEMENTARY INTRODUCTION TO MODERN CONVEX GEOMETRY

43

x

y

d(x, y)

Figure 26. The Euclidean metric on the sphere. A spherical cap (right) is a ball
for this metric.

enough (despite its difficulty): but it has a startling consequence.

Suppose

α =

1

2

, so that A has the measure of a hemisphere H. Then, for each positive ε,

the set A

ε

has measure at least that of the set H

ε

, illustrated in Figure 27. The

complement of H

ε

is a spherical cap that, as we saw in Lecture 2, has measure

about e

−nε

2

/2

. Hence σ(A

ε

)

1−e

−nε

2

/2

, so almost the entire sphere lies within

distance ε of A, even though there may be points rather far from A. The measure
and the metric on the sphere “don’t match”: the mass of σ concentrates very
close to any set of measure

1

2

. This is clearly related to the situation described

in Lecture 1, in which we found most of the mass of the ball concentrated near
each hyperplane: but now the phenomenon occurs for any set of measure

1

2

.

The phenomenon just described becomes even more striking when reinter-

preted in terms of Lipschitz functions. Suppose f : S

n−1

R

is a function on

the sphere that is 1-Lipschitz: that is, for any pair of points θ and φ on the
sphere,

|f(θ) − f(φ)| ≤ |θ − φ| .

There is at least one number M , the median of f , for which both the sets (f

≤ M)

and (f

≥ M) have measure at least

1

2

. If a point x has distance at most ε from

ε

Figure 27. An ε-neighbourhood of a hemisphere.

background image

44

KEITH BALL

(f

≤ M), then (since f is 1-Lipschitz)

f (x)

≤ M + ε.

By the isoperimetric inequality all but a tiny fraction of the points on the sphere
have this property:

σ(f > M + ε)

≤ e

−nε

2

/2

.

Similarly, f is larger than M

− ε on all but a fraction of the sphere. Putting

these statements together we get

σ(

|f − M| > ε) 2e

−nε

2

/2

.

So, although f may vary by as much as 2 between a point of the sphere and
its opposite, the function is nearly equal to M on almost the entire sphere: f is
practically constant.

In the case of the sphere we thus have the following pair of properties.

(i) If A

Ω with µ(A) =

1

2

then µ(A

ε

)

1 − e

−nε

2

/2

.

(ii) If f : Ω

R

is 1-Lipschitz there is a number M for which

µ(

|f − M| > ε) 2e

−nε

2

/2

.

Each of these statements may be called an approximate isoperimetric inequal-
ity. We have seen how the second can be deduced from the first. The reverse
implication also holds (apart from the precise constants involved). (To see why,
apply the second property to the function given by f (x) = d(x, A).)

In many applications, exact solutions of the isoperimetric problem are not as

important as deviation estimates of the kind we are discussing. In some cases
where the exact solutions are known, the two properties above are a good deal
easier to prove than the solutions: and in a great many situations, an exact
isoperimetric inequality is not known, but the two properties are.

The for-

mal similarity between property 2 and Bernstein’s inequality of the last lecture
is readily apparent. There are ways to make this similarity much more than
merely formal: there are deviation inequalities that have implications for Lips-
chitz functions and imply Bernstein’s inequality, but we shall not go into them
here.

In our second example, the space Ω will be

R

n

equipped with the ordinary

Euclidean distance.

The measure will be the standard Gaussian probability

measure on

R

n

with density

γ(x) = (2π)

−n/2

e

−|x|

2

/2

.

The solutions of the isoperimetric problem in Gauss space were found by Borell
[1975]. They are half-spaces. So, in particular, if A

R

n

and µ(A) =

1

2

, then

µ(A

ε

) is at least as large as µ(H

ε

), where H is the half-space

{x ∈

R

n

: x

1

0}

and so H

ε

=

{x : x

1

≤ ε}: see Figure 28.

background image

AN ELEMENTARY INTRODUCTION TO MODERN CONVEX GEOMETRY

45

ε

H

ε

Figure 28. An ε-neighbourhood of a half-space.

The complement of H

ε

has measure

1

2π

Z

ε

e

−t

2

/2

dt

≤ e

−ε

2

/2

.

Hence,

µ(A

ε

)

1 − e

−ε

2

/2

.

Since n does not appear in the exponent, this looks much weaker than the state-
ment for the sphere, but we shall see that the two are more or less equivalent.

Borell proved his inequality by using the isoperimetric inequality on the

sphere. A more direct proof of a deviation estimate like the one just derived
was found by Maurey and Pisier, and their argument gives a slightly stronger,
Sobolev-type inequality [Pisier 1989, Chapter 4]. We too shall aim directly for
a deviation estimate, but a little background to the proof may be useful.

There was an enormous growth in understanding of approximate isoperimetric

inequalities during the late 1980s, associated most especially with the name of
Talagrand. The reader whose interest has been piqued should certainly look at
Talagrand’s gorgeous articles [1988; 1991a], in which he describes an approach to
deviation inequalities in product spaces that involves astonishingly few structural
hypotheses. In a somewhat different vein (but prompted by his earlier work),
Talagrand [1991b] also found a general principle, strengthening the approximate
isoperimetric inequality in Gauss space. A simplification of this argument was
found by Maurey [1991]. The upshot is that a deviation inequality for Gauss
space can be proved with an extremely short argument that fits naturally into
these notes.

Theorem 8.1 (Approximate isoperimetric inequality for Gauss space).

Let A

R

n

be measurable and let µ be the standard Gaussian measure on

R

n

.

Then

Z

e

d(x,A)

2

/4

1

µ(A)

.

Consequently, if µ(A) =

1

2

,

µ(A

ε

)

1 2e

−ε

2

/4

.

background image

46

KEITH BALL

Proof.

We shall deduce the first assertion directly from the Pr´

ekopa–Leindler

inequality (with λ =

1

2

) of Lecture 5. To this end, define functions f , g, and m

on

R

n

, as follows:

f (x) = e

d(x,A)

2

/4

γ(x),

g(x) = χ

A

(x) γ(x),

m(x) = γ(x),

where γ is the Gaussian density. The assertion to be proved is that

Z

e

d(x,A)

2

/4



µ(A)

1,

which translates directly into the inequality

Z

R

n

f

 Z

R

n

g



Z

R

n

m



2

.

By the Pr´

ekopa–Leindler inequality it is enough to check that, for any x and y

in

R

n

,

f (x)g(y)

≤ m



x + y

2



2

.

It suffices to check this for y

∈ A, since otherwise g(y) = 0. But, in this case,

d(x, A)

≤ |x − y|. Hence

(2π)

n

f (x)g(y) = e

d(x,A)

2

/4

e

−x

2

/2

e

−y

2

/2

exp



|x − y|

2

4

|x|

2

2

|y|

2

2



= exp



|x + y|

2

4



=



exp



1

2

x + y

2

2



2

= (2π)

n

m



x + y

2



2

,

which is what we need.

To deduce the second assertion from the first, we use Markov’s inequality, very

much as in the proof of Bernstein’s inequality of the last lecture. If µ(A) =

1

2

,

then

Z

e

d(x,A)

2

/4

2.

The integral is at least

e

ε

2

/4

µ d(x, A)

≥ ε



.

So

µ d(x, A)

≥ ε



2e

−ε

2

/4

,

and the assertion follows.



background image

AN ELEMENTARY INTRODUCTION TO MODERN CONVEX GEOMETRY

47

It was mentioned earlier that the Gaussian deviation estimate above is essentially
equivalent to the concentration of measure on S

n−1

. This equivalence depends

upon the fact that the Gaussian measure in

R

n

is concentrated in a spherical shell

of thickness approximately 1, and radius approximately

n. (Recall that the

Euclidean ball of volume 1 has radius approximately

n.) This concentration is

easily checked by direct computation using integration in spherical polars: but
the inequality we just proved will do the job instead. There is an Euclidean
ball of some radius R whose Gaussian measure is

1

2

. According to the theorem

above, Gaussian measure concentrates near the boundary of this ball. It is not
hard to check that R is about

n. This makes it quite easy to show that the

deviation estimate for Gaussian measure guarantees a deviation estimate on the
sphere of radius

n with a decay rate of about e

−ε

2

/4

. If everything is scaled

down by a factor of

n, onto the sphere of radius 1, we get a deviation estimate

that decays like e

−nε

2

/4

and n now appears in the exponent. The details are left

to the reader.

The reader will probably have noticed that these estimates for Gauss space

and the sphere are not quite as strong as those advertised earlier, because in
each case the exponent is . . . ε

2

/4 . . . instead of . . . ε

2

/2 . . .. In some applications,

the sharper results are important, but for our purposes the difference will be
irrelevant. It was pointed out to me by Talagrand that one can get as close
as one wishes to the correct exponent . . . ε

2

/2 . . . by using the Pr´

ekopa–Leindler

inequality with λ close to 1 instead of

1

2

and applying it to slightly different f

and g.

For the purposes of the next lecture we shall assume an estimate of e

−ε

2

/2

,

even though we proved a weaker estimate.

Lecture 9. Dvoretzky’s Theorem

Although this is the ninth lecture, its subject, Dvoretzky’s Theorem, was re-

ally the start of the modern theory of convex geometry in high dimensions. The
phrase “Dvoretzky’s Theorem” has become a generic term for statements to the
effect that high-dimensional bodies have almost ellipsoidal slices. Dvoretzky’s
original proof shows that any symmetric convex body in

R

n

has almost ellip-

soidal sections of dimension about

log n. A few years after the publication

of Dvoretzky’s work, Milman [Milman 1971] found a very different proof, based
upon the concentration of measure, which gave slices of dimension log n. As we
saw in Lecture 2 this is the best one can get in general. Milman’s argument gives
the following.

Theorem 9.1.

There is a positive number c such that , for every ε > 0 and

every natural number n, every symmetric convex body of dimension n has a slice
of dimension

k

2

log(1 + ε

1

)

log n

background image

48

KEITH BALL

that is within distance 1 + ε of the k-dimensional Euclidean ball .

There have been many other proofs of similar results over the years. A par-
ticularly elegant one [Gordon 1985] gives the estimate k

≥ cε

2

log n (removing

the logarithmic factor in ε

1

), and this estimate is essentially best possible. We

chose to describe Milman’s proof because it is conceptually easier to motivate
and because the concentration of measure has many other uses. A few years ago,
Schechtman found a way to eliminate the log factor within this approach, but
we shall not introduce this subtlety here. We shall also not make any effort to
be precise about the dependence upon ε.

With the material of Lecture 8 at our disposal, the plan of proof of Theorem

9.1 is easy to describe. We start with a symmetric convex body and we consider
a linear image K whose maximal volume ellipsoid is the Euclidean ball. For this
K we will try to find almost spherical sections, rather than merely ellipsoidal
ones. Let

k · k be the norm on

R

n

whose unit ball is K. We are looking for a

k-dimensional space H with the property that the function

θ

7→ kθk

is almost constant on the Euclidean sphere of H, H

∩ S

n−1

. Since K contains

B

n

2

, we have

kxk ≤ |x| for all x ∈

R

n

, so for any θ and φ in S

n−1

,

|kθk − kφk| ≤ kθ − φk ≤ |θ − φ|.

Thus

k · k is a Lipschitz function on the sphere in

R

n

, (indeed on all of

R

n

). (We

used the same idea in Lecture 4.) From Lecture 8 we conclude that the value of
kθk is almost constant on a very large proportion of S

n−1

: it is almost equal to

its average

M =

Z

S

n−1

kθk dσ,

on most of S

n−1

.

We now choose our k-dimensional subspace at random. (The exact way to do

this will be described below.) We can view this as a random embedding

T :

R

k

R

n

.

For any particular unit vector ψ

R

k

, there is a very high probability that its

image T ψ will have norm

kT ψk close to M. This means that even if we select

quite a number of vectors ψ

1

, ψ

2

, . . . , ψ

m

in S

k−1

we can guarantee that there

will be some choice of T for which all the norms

kT ψ

i

k will be close to M. We

will thus have managed to pin down the radius of our slice in many different
directions. If we are careful to distribute these directions well over the sphere in

R

k

, we may hope that the radius will be almost constant on the entire sphere.

For these purposes, “well distributed” will mean that all points of the sphere in

R

k

are close to one of our chosen directions. As in Lecture 2 we say that a set

1

, ψ

2

, . . . , ψ

m

} in S

k−1

is a δ-net for the sphere if every point of S

k−1

is within

background image

AN ELEMENTARY INTRODUCTION TO MODERN CONVEX GEOMETRY

49

(Euclidean) distance δ of at least one ψ

i

. The arguments in Lecture 2 show that

S

k−1

has a δ-net with no more than

m =



4

δ



k

elements. The following lemma states that, indeed, pinning down the norm on
a very fine net, pins it down everywhere.

Lemma 9.2.

Let

k · k be a norm on

R

k

and suppose that for each point ψ of

some δ-net on S

k−1

, we have

M (1

− γ) ≤ kψk ≤ M(1 + γ)

for some γ > 0. Then, for every θ

∈ S

k−1

,

M (1

− γ − 2δ)

1

− δ

≤ kθk ≤

M (1 + γ)

1

− δ

.

Proof.

Clearly the value of M plays no real role here so assume it is 1. We

start with the upper bound. Let C be the maximum possible ratio

kxk/|x| for

nonzero x and let θ be a point of S

k−1

with

kθk = C. Choose ψ in the δ-net

with

|θ − ψ| ≤ δ. Then kθ − ψk ≤ C|θ − ψ| ≤ Cδ, so

C =

kθk ≤ kψk + kθ − ψk ≤ (1 + γ) + Cδ.

Hence

C

(1 + γ)

1

− δ

.

To get the lower bound, pick some θ in the sphere and some ψ in the δ-net

with

|ψ − θ| ≤ δ. Then

(1

− γ) ≤ kψk ≤ kθk + kψ − θk ≤ kθk +

(1 + γ)

1

− δ

|ψ − θ| ≤ kθk +

(1 + γ)δ

1

− δ

.

Hence

kθk ≥



1

− γ −

δ(1 + γ)

1

− δ



=

(1

− γ − 2δ)

1

− δ

.



According to the lemma, our approach will give us a slice that is within distance

1 + γ

1

− γ − 2δ

of the Euclidean ball (provided we satisfy the hypotheses), and this distance can
be made as close as we wish to 1 if γ and δ are small enough.

We are now in a position to prove the basic estimate.

Theorem 9.3.

Let K be a symmetric convex body in

R

n

whose ellipsoid of

maximal volume is B

n

2

and put

M =

Z

S

n−1

kθk dσ

background image

50

KEITH BALL

as above. Then K has almost spherical slices whose dimension is of the order of
nM

2

.

Proof.

Choose γ and δ small enough to give the desired accuracy, in accordance

with the lemma.

Since the function θ

7→ kθk is Lipschitz (with constant 1) on S

n−1

, we know

from Lecture 8 that, for any t

0,

σ

k

θ

k − M

> t



2e

−nt

2

/2

.

In particular,

σ

k

θ

k − M

> M γ



2e

−nM

2

γ

2

/2

.

So

M (1

− γ) ≤ kθk ≤ M(1 + γ)

on all but a proportion 2e

−nM

2

γ

2

/2

of the sphere.

Let

A be a δ-net on the sphere in

R

k

with at most (4)

k

elements. Choose

a random embedding of

R

k

in

R

n

: more precisely, fix a particular copy of

R

k

in

R

n

and consider its images under orthogonal transformations U of

R

n

as

a random subspace with respect to the invariant probability on the group of
orthogonal transformations. For each fixed ψ in the sphere of

R

k

, its images

U ψ, are uniformly distributed on the sphere in

R

n

. So for each ψ

∈ A, the

inequality

M (1

− γ) ≤ kUψk ≤ M(1 + γ)

holds for U outside a set of measure at most 2e

−nM

2

γ

2

/2

. So there will be at

least one U for which this inequality holds for all ψ in

A, as long as the sum of

the probabilities of the bad sets is at most 1. This is guaranteed if



4

δ



k

2e

−nM

2

γ

2

/2

< 1.

This inequality is satisfied by k of the order of

nM

2

γ

2

2 log(4)

.



Theorem 9.3 guarantees the existence of spherical slices of K of large dimension,
provided the average

M =

Z

S

n−1

kθk dσ

is not too small. Notice that we certainly have M

1 since kxk ≤ |x| for all x.

In order to get Theorem 9.1 from Theorem 9.3 we need to get a lower estimate
for M of the order of

log n

n

.

background image

AN ELEMENTARY INTRODUCTION TO MODERN CONVEX GEOMETRY

51

This is where we must use the fact that B

n

2

is the maximal volume ellipsoid in

K. We saw in Lecture 3 that in this situation K

nB

n

2

, so

kxk ≥ |x|/

n for

all x, and hence

M

1

n

.

But this estimate is useless, since it would not give slices of dimension bigger
than 1. It is vital that we use the more detailed information provided by John’s
Theorem.

Before we explain how this works, let’s look at our favourite examples. For

specific norms it is usually much easier to compute the mean M by writing it as
an integral with respect to Gaussian measure on

R

n

. As in Lecture 8 let µ be

the standard Gaussian measure on

R

n

, with density

(2π)

−n/2

e

−|x|

2

/2

.

By using polar coordinates we can write

Z

S

n−1

kθk dσ =

Γ(n/2)

2Γ((n + 1)/2)

Z

R

n

kxk dµ(x) >

1

n

Z

R

n

kxk dµ(x).

The simplest norm for which to calculate is the `

1

norm. Since the body we

consider is supposed to have B

n

2

as its maximal ellipsoid we must use

nB

n

1

, for

which the corresponding norm is

kxk =

1

n

n

X

1

|x

i

|.

Since the integral of this sum is just n times the integral of any one coordinate
it is easy to check that

1

n

Z

R

n

kxk dµ(x) =

r

2

π

.

So for the scaled copies of B

n

1

, we have M bounded below by a fixed number, and

Theorem 9.3 guarantees almost spherical sections of dimension proportional to
n. This was first proved, using exactly the method described here, in [Figiel et al.
1977], which had a tremendous influence on subsequent developments. Notice
that this result and Kaˇsin’s Theorem from Lecture 4 are very much in the same
spirit, but neither implies the other. The method used here does not achieve
dimensions as high as n/2 even if we are prepared to allow quite a large distance
from the Euclidean ball. On the other hand, the volume-ratio argument does
not give sections that are very close to Euclidean: the volume ratio is the closest
one gets this way. Some time after Kaˇsin’s article appeared, the gap between
these results was completely filled by Garnaev and Gluskin [Garnaev and Gluskin
1984]. An analogous gap in the general setting of Theorem 9.1, namely that the
existing proofs could not give a dimension larger than some fixed multiple of
log n, was recently filled by Milman and Schechtman.

background image

52

KEITH BALL

What about the cube? This body has B

n

2

as its maximal ellipsoid, so our job

is to estimate

1

n

Z

R

n

max

|x

i

| dµ(x).

At first sight this looks much more complicated than the calculation for B

n

1

,

since we cannot simplify the integral of a maximum. But, instead of estimating
the mean of the function max

|x

i

|, we can estimate its median (and from Lecture

8 we know that they are not far apart). So let R be the number for which

µ (max

|x

i

| ≤ R) = µ (max |x

i

| ≥ R) =

1

2

.

From the second identity we get

1

n

Z

R

n

max

|x

i

| dµ(x)

R

2

n

.

We estimate R from the first identity. It says that the cube [

−R, R]

n

has Gauss-

ian measure

1

2

. But the cube is a “product” so

µ ([

−R, R]

n

) =

1

2π

Z

R

−R

e

−t

2

/2

dt

!

n

.

In order for this to be equal to

1

2

we need the expression

1

2π

Z

R

−R

e

−t

2

/2

dt

to be about 1

(log 2)/n. Since the expression approaches 1 roughly like

1

− e

−R

2

/2

,

we get an estimate for R of the order of

log n. From Theorem 9.3 we then

recover the simple result of Lecture 2 that the cube has almost spherical sections
of dimension about log n.

There are many other bodies and classes of bodies for which M can be ef-

ficiently estimated. For example, the correct order of the largest dimension of
Euclidean slice of the `

n

p

balls, was also computed in the paper [Figiel et al. 1977]

mentioned earlier.

We would like to know that for a general body with maximal ellipsoid B

n

2

we

have

Z

R

n

kxk dµ(x) (constant)

p

log n

(9.1)

just as we do for the cube. The usual proof of this goes via the Dvoretzky–Rogers
Lemma, which can be proved using John’s Theorem. This is done for example in
[Pisier 1989]. Roughly speaking, the Dvoretzky–Rogers Lemma builds something
like a cube around K, at least in a subspace of dimension about

n

2

, to which we

then apply the result for the cube. However, I cannot resist mentioning that the
methods of Lecture 6, involving sharp convolution inequalities, can be used to

background image

AN ELEMENTARY INTRODUCTION TO MODERN CONVEX GEOMETRY

53

show that among all symmetric bodies K with maximal ellipsoid B

n

2

the cube

is precisely the one for which the integral in (9.1) is smallest. This is proved by
showing that for each r, the Gaussian measure of rK is at most that of [

−r, r]

n

.

The details are left to the reader.

This last lecture has described work that dates back to the seventies. Although

some of the material in earlier lectures is more recent (and some is much older),
I have really only scratched the surface of what has been done in the last twenty
years. The book of Pisier to which I have referred several times gives a more
comprehensive account of many of the developments. I hope that readers of
these notes may feel motivated to discover more.

Acknowledgements

I would like to thank Silvio Levy for his help in the preparation of these notes,

and one of the workshop participants, John Mount, for proofreading the notes
and suggesting several improvements. Finally, a very big thank you to my wife
Sachiko Kusukawa for her great patience and constant love.

References

[Ball 1990] K. M. Ball, “Volumes of sections of cubes and related problems”, pp. 251–

260 in Geometric aspects of functional analysis (Israel Seminar, 1987–1988), edited
by J. Lindenstrauss and V. D. Milman, Lecture Notes in Math.

1376, Springer,

1990.

[Ball 1991]

K. M. Ball, “Volume ratios and a reverse isoperimetric inequality”, J.

London Math. Soc.

44 (1991), 351–359.

[Beckner 1975]

W. Beckner, “Inequalities in Fourier analysis”, Ann. of Math.

102

(1975), 159–182.

[Borell 1975] C. Borell, “The Brunn–Minkowski inequality in Gauss space”, Inventiones

Math.

30 (1975), 205–216.

[Bourgain and Milman 1987] J. Bourgain and V. Milman, “New volume ratio properties

for convex symmetric bodies in

R

n

”, Invent. Math.

88 (1987), 319–340.

[Bourgain et al. 1989] J. Bourgain, J. Lindenstrauss, and V. Milman, “Approximation

of zonoids by zonotopes”, Acta Math.

162 (1989), 73–141.

[Brascamp and Lieb 1976a]

H. J. Brascamp and E. H. Lieb, “Best constants in

Young’s inequality, its converse and its generalization to more than three functions”,
Advances in Math.

20 (1976), 151–173.

[Brascamp and Lieb 1976b] H. J. Brascamp and E. H. Lieb, “On extensions of the

Brunn–Minkowski and Pr´

ekopa–Leindler theorems, including inequalities for log

concave functions, and with an application to the diffusion equation”, J. Funct.
Anal.

22 (1976), 366–389.

[Brøndsted 1983] A. Brøndsted, An introduction to convex polytopes, Graduate Texts

in Math.

90, Springer, New York, 1983.

background image

54

KEITH BALL

[Busemann 1949] H. Busemann, “A theorem on convex bodies of the Brunn–Minkowski

type”, Proc. Nat. Acad. Sci. USA

35 (1949), 27–31.

[Figiel et al. 1977]

T. Figiel, J. Lindenstrauss, and V. Milman, “The dimension of

almost spherical sections of convex bodies”, Acta Math.

139 (1977), 53–94.

[Garnaev and Gluskin 1984] A. Garnaev and E. Gluskin, “The widths of a Euclidean

ball”, Dokl. A. N. USSR

277 (1984), 1048–1052. In Russian.

[Gordon 1985] Y. Gordon, “Some inequalities for Gaussian processes and applications”,

Israel J. Math.

50 (1985), 265–289.

[Hoeffding 1963] W. Hoeffding, “Probability inequalities for sums of bounded random

variables”, J. Amer. Statist. Assoc.

58 (1963), 13–30.

[John 1948] F. John, “Extremum problems with inequalities as subsidiary conditions”,

pp. 187–204 in Studies and essays presented to R. Courant on his 60th birthday (Jan.
8, 1948), Interscience, New York, 1948.

[Johnson and Schechtman 1982] W. B. Johnson and G. Schechtman, “Embedding `

m

p

into `

n

1

”, Acta Math.

149 (1982), 71–85.

[Kaˇ

sin 1977] B. S. Kaˇ

sin, “The widths of certain finite-dimensional sets and classes of

smooth functions”, Izv. Akad. Nauk SSSR Ser. Mat.

41:2 (1977), 334–351, 478. In

Russian.

[Lieb 1990] E. H. Lieb, “Gaussian kernels have only Gaussian maximizers”, Invent.

Math.

102 (1990), 179–208.

[Lıusternik 1935]

L. A. Lıusternik, “Die Brunn–Minkowskische Ungleichung f¨

ur

beliebige messbare Mengen”, C. R. Acad. Sci. URSS

8 (1935), 55–58.

[Maurey 1991]

B. Maurey, “Some deviation inequalities”, Geom. Funct. Anal.

1:2

(1991), 188–197.

[Milman 1971] V. Milman, “A new proof of A. Dvoretzky’s theorem on cross-sections

of convex bodies”, Funkcional. Anal. i Priloˇ

zen

5 (1971), 28–37. In Russian.

[Milman 1985] V. Milman, “Almost Euclidean quotient spaces of subspaces of finite

dimensional normed spaces”, Proc. Amer. Math. Soc.

94 (1985), 445–449.

[Pisier 1989] G. Pisier, The volume of convex bodies and Banach space geometry, Tracts

in Math.

94, Cambridge U. Press, Cambridge, 1989.

[Rogers 1964] C. A. Rogers, Packing and covering, Cambridge U. Press, Cambridge,

1964.

[Schneider 1993] R. Schneider, Convex bodies: the Brunn–Minkowski theory, Encyclo-

pedia of Math. and its Applications

44, Cambridge U. Press, 1993.

[Szarek 1978] S. J. Szarek, “On Kashin’s almost Euclidean orthogonal decomposition

of `

n

1

”, Bull. Acad. Polon. Sci. S´

er. Sci. Math. Astronom. Phys.

26 (1978), 691–694.

[Talagrand 1988]

M. Talagrand, “An isoperimetric inequality on the cube and the

Khintchine-Kahane inequalities”, Proc. Amer. Math. Soc.

104 (1988), 905–909.

[Talagrand 1990] M. Talagrand, “Embedding subspaces of L

1

into `

N

1

”, Proc. Amer.

Math. Soc.

108 (1990), 363–369.

[Talagrand 1991a] M. Talagrand, “A new isoperimetric inequality”, Geom. Funct. Anal.

1:2 (1991), 211–223.

background image

AN ELEMENTARY INTRODUCTION TO MODERN CONVEX GEOMETRY

55

[Talagrand 1991b]

M. Talagrand, “A new isoperimetric inequality and the concen-

tration of measure phenomenon”, pp. 94–124 in Geometric aspects of functional
analysis
(Israel Seminar, 1989–1990), edited by J. Lindenstrauss and V. D. Milman,
Lecture Notes in Math.

1469, Springer, 1991.

[Tomczak-Jaegermann 1988] N. Tomczak-Jaegermann, Banach–Mazur distances and

finite-dimensional operator ideals, Pitman Monographs and Surveys in Pure and
Applied Math.

38, Longman, Harlow, 1988.

[Williams 1991] D. Williams, Probability with martingales, Cambridge Mathematical

Textbooks, Cambridge University Press, Cambridge, 1991.

Index

affine transformation, 8, 13, 32
almost, see approximate
AM/GM inequality, 17, 19, 29, 31
approximate

ball, 8, 20

intersection, 20
section, 8, 9, 19, 24, 49

ellipsoid, see approximate ball
isoperimetric inequality, 44, 45

area, see also volume
arithmetic mean, see AM/GM inequality
average radius, 7
Azuma, Kazuoki, 39
Azuma–Hoeffding inequality, 39

B

n

1

, 3

B

n

2

, 4

ball, see also approximate ball

and cube, 2, 8, 9, 19
and octahedron, 3
and simplex, 2
central section, 6
circumscribed, 2, 3, 8
distribution of volume, 6
Euclidean, 4
for arbitrary norm, 8
inscribed, 2, 3, 8
near polytope, 9
section, 6
similarities to convex body, 2
volume, 4

Ball, Keith M., 25, 33, 36
Beckner, William, 34, 37
Bernoulli distribution, 39
Bernstein’s inequality, 39, 41, 44, 46
Bernstein, Sergei N., 39
Bollob´

as, B´

ela, 1

bootstrapping of probabilities, 24
Borell, Christer, 44
bound, see inequality
Bourgain, Jean, 25
Brascamp, Herm Jan, 30, 34, 35, 37
Brascamp–Lieb reverse inequality, 37
Brunn’s Theorem, 26, 28, 29
Brunn, Hermann, 25, 26
Brunn–Minkowski

inequality, 25, 28, 32, 41

functional, 29
multiplicative, 29

theory, 2

Brøndsted, Arne, 2
Busemann’s Theorem, 32
Busemann, Herbert, 32

cap, 10, 22

radius, 11
volume, 10, 11

Cauchy–Schwarz inequality, 24
central

limit theorem, 37, 39
section, 6

centrally symmetric, see symmetric
characteristic function, 29
Chebyshev’s inequality, 40
chordal metric, 42
circumscribed ball, 2, 8
classical convex geometry, 2
coin toss, 39
combinatorial

theory of polytopes, 2

combinatorics, 32
concave function, 25, 27, 29
concentration, see also distribution

of measure, 7, 41, 47

background image

56

KEITH BALL

cone, 3, 26

spherical, 12
volume, 3

contact point, 13
convex

body

definition, 2
ellipsoids inside, 13
similarities to ball, 2
symmetric, 8

hull, 2, 3

convolution

generalised, 34
inequalities, 52

cotype-2 property, 25
cross-polytope, 3, 8

and ball, 13
volume, 3

ratio, 19, 21

cube, 2, 7, 33, 52, 53

and ball, 8, 9, 13, 19
distribution of volume, 41
maximal ellipsoid, 15
sections of, 8
volume, 7

distribution, 7
ratio, 19, 21

deck of cards, 25
density, see distribution
determinant maximization, 19
deviation estimate, 24, 39

geometric, 41
in Gauss space, 45

differential equations, 2
dimension-independent constant, 6, 20,

25, 47

distance

between convex bodies, 8

distribution, see also concentration, see

also volume distribution

Gaussian, 6, 12
normal, 6
of volume, 12

duality, 17, 19
Dvoretzky’s Theorem, 47
Dvoretzky, Aryeh, 47, 52
Dvoretzky–Rogers Lemma, 52

ellipsoid, 13

maximal, see maximal ellipsoid

minimal, 37
polar, 17

Euclidean

metric, 11

on sphere, 42

norm, 2, 14

expectation, 37

Figiel, Tadeusz, 51, 52
Fourier transform, 37
Fubini’s Theorem, 31, 38
functional, 2, 18

analysis, 2
Brunn–Minkowski inequality, 29

Garnaev, Andre˘ı, 51
Gaussian

distribution, 6, 12, 34, 36, 39, 44, 46, 47,

51, 53

measure, see Gaussian distribution

generalized convolution, 34
geometric

mean, see AM/GM inequality

geometry

of numbers, 2

Gluskin, E. D., 51
Gordon, Yehoram, 48

older inequality, 30

Hadamard matrix, 24
Hahn–Banach separation theorem, 2
harmonic

analysis, 34, 37

Hoeffding, Wassily, 39
homogeneity, 30

identity

resolution of, 14, 19

independence

from

dimension,

see

dimension-independent constant

independent variables, 37
inequality

AM/GM, 17, 19, 29, 31
Bernstein’s, 39, 41, 44, 46
Brunn–Minkowski, 25, 28, 32, 41

functional, 29
multiplicative, 29

Cauchy–Schwarz, 24
Chebyshev’s, 40
convolution, 52
deviation, see deviation estimate
for cap volume, 11
for norms of convolutions, 34

background image

AN ELEMENTARY INTRODUCTION TO MODERN CONVEX GEOMETRY

57

older, 30

isoperimetric, 28, 32, 41, 44, 45
Markov’s, 40, 46
Pr´

ekopa–Leindler, 30, 46, 47

Young’s, 34

inertia tensor, 14
information theory, 2, 32
inscribed ball, 2, 8
intersection

of convex bodies, 20

invariant measure on O(n), 22, 50
isoperimetric

inequality, 28, 32, 41

approximate, 44, 45
in Gauss space, 45
on sphere, 42

problem

in Gauss space, 44

John’s Theorem, 13, 33, 36, 51, 52

extensions, 19
proof of, 16

John, Fritz, 13
Johnson, William B., 25

Kaˇ

sin’s Theorem, 20, 51

Kaˇ

sin, B. S., 20

Kusukawa, Sachiko, 53

L

p

norm, 34

`

1

norm, 3, 8, 24, 51

evy, Paul, 42

large deviation inequalities, see deviation

estimate

Leindler, L´

aszl´

o, 30

Levy, Silvio, 53
Lieb, Elliott H., 30, 34, 35, 37
Lindenstrauss, Joram, 1, 25, 51, 52
linear

programming, 2

Lipschitz function, 43, 50
Lyusternik, Lazar A., 28

Markov’s inequality, 40, 46
mass, see also volume

distribution on contact points, 14

Maurey, Bernard, 45
maximal

ellipsoid, 13, 21, 34, 48, 49, 51, 53

for cube, 15

mean, see AM/GM inequality
metric

and measure, 42
on space of convex bodies, 8
on sphere, 11

Milman, Vitali D., 25, 41, 47, 51, 52
minimal ellipsoid, 19, 37
Minkowski

addition, 28, 42

Minkowski, Hermann, 27
Mount, John, 53
MSRI, 1
multiplicative

Brunn–Minkowski inequality, 29

N(x), 22
net, 11, 48
nonsymmetric convex body, 13, 15
norm

`

1

, 3, 8, 24, 51

and radius, 8
Euclidean, 2, 14, 21
of convolution, 34
with given ball, 8, 21, 48

normal distribution, 6
normalisation of integral, 5, 39

octahedron, see cross-polytope
omissions, 2
orthant, 3
orthogonal

group, 22, 50
projection, 14

orthonormal basis, 13, 14

partial differential equations, 2
Pisier, Gilles, 20, 45, 52
polar

coordinates, 4, 7, 47, 51
ellipsoid, 17

polytope, 8

as section of cube, 9
combinatorial theory of –s, 2
near ball, 9

Pr´

ekopa, Andr´

as, 30

Pr´

ekopa–Leindler inequality, 30, 46, 47

probability

bootstrapping of, 24
measure, 37, 42
that. . . , 37
theory, 2, 24, 37

projection

orthogonal, 14

QS-Theorem, 25

background image

58

KEITH BALL

quotient of a subspace, 25

r(θ), 7
radius

and norm, 8
and volume for ball, 5
average, 7
in a direction, 7, 32
of body in a direction, 21

Ramsey theory, 24
random

subspace, 48, 50
variable, 37

ratio

between volumes, see volume ratio

resolution of the identity, 14, 19
reverse

Brascamp–Lieb inequality, 37

rigidity condition, 16
Rogers, Claude Ambrose, 9, 52

S

n−1

, 4

Schechtman, Gideon, 1, 25, 48, 51
Schneider, Rolf, 2, 32
section

ellipsoidal, see approximate ball
length of, 25
of ball, 6
of cube, 8, 9
parallel, 27
spherical, see approximate ball

sections

almost spherical, 50

1-separated set, 12
separation

theorem, 18

shaking down a set, 25
σ, 5
σ-field, 38

simplex, 15, 33

regular, 2

slice, see section
sphere, see also ball

cover by caps, 10
measure on, 5
volume, 5

spherical

cap, see cap
cone, 12
coordinates, see polar coordinates
section, see approximate ball

Stirling’s formula, 5, 6
supporting hyperplane, 2, 16
symmetric body, 8, 15, 25, 33, 47, 49
Szarek, Stanis law Jerzy, 20, 21

Talagrand, Michel, 25, 45
tensor product, 14
Tomczak-Jaegermann, Nicole, 19

vol, 2
volume

distribution

of ball, 6
of cube, 7, 41

of ball, 4
of cone, 3
of cross-polytope, 3
of ellipsoid, 13
of symmetric body, 8
ratio, 19, 21, 25, 33, 36, 51

vr, 21

Walsh matrix, 24
weighted average, 14, 27
Williams, David, 39

Young’s inequality, 34

Keith Ball

Department of Mathematics

University College

University of London

London

United Kingdom

kmb@math.ucl.ac.uk


Wyszukiwarka

Podobne podstrony:
Gudmundsson S An introduction to Riemannian geometry(web draft, 2006)(94s) MDdg
Zizek, Slavoj Looking Awry An Introduction to Jacques Lacan through Popular Culture
An Introduction to the Kabalah
An Introduction to USA 6 ?ucation
An Introduction to Database Systems, 8th Edition, C J Date
PP 3 4 INTRODUCTION TO DRAMA Modern Drama
An Introduction to Extreme Programming
Khudaverdian Introduction to Geometry
Adler M An Introduction to Complex Analysis for Engineers
An Introduction to American Lit Nieznany (2)
(ebook pdf) Mathematics An Introduction To Cryptography ZHS4DOP7XBQZEANJ6WTOWXZIZT5FZDV5FY6XN5Q
An Introduction to USA 1 The Land and People
An Introduction to USA 4 The?onomy and Welfare
An Introduction to USA 7 American Culture and Arts
An Introduction To Olap In Sql Server 2005
An Introduction to Yang Mills Theory
Introduction to Descriptive Geometry
An Introduction to Linux Systems?ministration

więcej podobnych podstron